[PATCH] uml: use host AIO support
authorJeff Dike <jdike@addtoit.com>
Sat, 3 Sep 2005 22:57:45 +0000 (15:57 -0700)
committerLinus Torvalds <torvalds@evo.osdl.org>
Mon, 5 Sep 2005 07:06:23 +0000 (00:06 -0700)
This patch makes UML use host AIO support when it (and
/usr/include/linux/aio_abi.h) are present.  This is only the support, with no
consumers - a consumer is coming in the next patch.

Signed-off-by: Jeff Dike <jdike@addtoit.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
arch/um/include/aio.h [new file with mode: 0644]
arch/um/include/init.h
arch/um/include/irq_kern.h
arch/um/kernel/irq.c
arch/um/os-Linux/Makefile
arch/um/os-Linux/aio.c [new file with mode: 0644]

diff --git a/arch/um/include/aio.h b/arch/um/include/aio.h
new file mode 100644 (file)
index 0000000..423bae9
--- /dev/null
@@ -0,0 +1,28 @@
+/*
+ * Copyright (C) 2004 Jeff Dike (jdike@karaya.com)
+ * Licensed under the GPL
+ */
+
+#ifndef AIO_H__
+#define AIO_H__
+
+enum aio_type { AIO_READ, AIO_WRITE, AIO_MMAP };
+
+struct aio_thread_reply {
+       void *data;
+       int err;
+};
+
+struct aio_context {
+       int reply_fd;
+       struct aio_context *next;
+};
+
+#define INIT_AIO_CONTEXT { .reply_fd   = -1, \
+                          .next        = NULL }
+
+extern int submit_aio(enum aio_type type, int fd, char *buf, int len,
+                     unsigned long long offset, int reply_fd,
+                      struct aio_context *aio);
+
+#endif
index 55c2693f877827abab61514a648c8ee20ce7d98b..cbd79a8d213dc4532b8f760d3b1ca207fa98b3f1 100644 (file)
@@ -111,7 +111,15 @@ extern struct uml_param __uml_setup_start, __uml_setup_end;
 
 #ifndef __KERNEL__
 
-#define __initcall(fn) static initcall_t __initcall_##fn __init_call = fn
+#define __define_initcall(level,fn) \
+       static initcall_t __initcall_##fn __attribute_used__ \
+       __attribute__((__section__(".initcall" level ".init"))) = fn
+
+/* Userspace initcalls shouldn't depend on anything in the kernel, so we'll
+ * make them run first.
+ */
+#define __initcall(fn) __define_initcall("1", fn)
+
 #define __exitcall(fn) static exitcall_t __exitcall_##fn __exit_call = fn
 
 #define __init_call __attribute__ ((unused,__section__ (".initcall.init")))
index 3af52a634c4c2c5b9d8eb16cd93b1cd879e91e6d..c222d56b14940192c9a1b8eb659e8c966a4fb656 100644 (file)
@@ -7,12 +7,15 @@
 #define __IRQ_KERN_H__
 
 #include "linux/interrupt.h"
+#include "asm/ptrace.h"
 
 extern int um_request_irq(unsigned int irq, int fd, int type,
                          irqreturn_t (*handler)(int, void *,
                                                 struct pt_regs *),
                          unsigned long irqflags,  const char * devname,
                          void *dev_id);
+extern int init_aio_irq(int irq, char *name,
+                       irqreturn_t (*handler)(int, void *, struct pt_regs *));
 
 #endif
 
index 9f18061ef4c98d16915f7f040709340ed6740ad2..dcd814971995ca858c2ce01667800437c65f43b9 100644 (file)
@@ -31,7 +31,7 @@
 #include "kern_util.h"
 #include "irq_user.h"
 #include "irq_kern.h"
-
+#include "os.h"
 
 /*
  * Generic, controller-independent functions:
@@ -168,13 +168,32 @@ void __init init_IRQ(void)
        }
 }
 
-/*
- * Overrides for Emacs so that we follow Linus's tabbing style.
- * Emacs will notice this stuff at the end of the file and automatically
- * adjust the settings for this buffer only.  This must remain at the end
- * of the file.
- * ---------------------------------------------------------------------------
- * Local variables:
- * c-file-style: "linux"
- * End:
- */
+int init_aio_irq(int irq, char *name, irqreturn_t (*handler)(int, void *,
+                                                            struct pt_regs *))
+{
+       int fds[2], err;
+
+       err = os_pipe(fds, 1, 1);
+       if(err){
+               printk("init_aio_irq - os_pipe failed, err = %d\n", -err);
+               goto out;
+       }
+
+       err = um_request_irq(irq, fds[0], IRQ_READ, handler,
+                            SA_INTERRUPT | SA_SAMPLE_RANDOM, name,
+                            (void *) (long) fds[0]);
+       if(err){
+               printk("init_aio_irq - : um_request_irq failed, err = %d\n",
+                      err);
+               goto out_close;
+       }
+
+       err = fds[1];
+       goto out;
+
+ out_close:
+       os_close_file(fds[0]);
+       os_close_file(fds[1]);
+ out:
+       return(err);
+}
index 4ddf540284ce601b66f54f089656846d4522ec80..351d96934679bd8def63e604d3d2b0315f7fb03e 100644 (file)
@@ -3,11 +3,15 @@
 # Licensed under the GPL
 #
 
-obj-y = elf_aux.o file.o process.o signal.o time.o tty.o user_syms.o drivers/ \
-       sys-$(SUBARCH)/
+obj-y = aio.o elf_aux.o file.o process.o signal.o time.o tty.o user_syms.o \
+       drivers/ sys-$(SUBARCH)/
 
-USER_OBJS := elf_aux.o file.o process.o signal.o time.o tty.o
+USER_OBJS := aio.o elf_aux.o file.o process.o signal.o time.o tty.o
 
 CFLAGS_user_syms.o += -DSUBARCH_$(SUBARCH)
 
+HAVE_AIO_ABI := $(shell [ -r /usr/include/linux/aio_abi.h ] && \
+       echo -DHAVE_AIO_ABI )
+CFLAGS_aio.o += $(HAVE_AIO_ABI)
+
 include arch/um/scripts/Makefile.rules
diff --git a/arch/um/os-Linux/aio.c b/arch/um/os-Linux/aio.c
new file mode 100644 (file)
index 0000000..f2ca299
--- /dev/null
@@ -0,0 +1,398 @@
+/*
+ * Copyright (C) 2004 Jeff Dike (jdike@addtoit.com)
+ * Licensed under the GPL
+ */
+
+#include <stdlib.h>
+#include <unistd.h>
+#include <signal.h>
+#include <errno.h>
+#include <sched.h>
+#include <sys/syscall.h>
+#include "os.h"
+#include "helper.h"
+#include "aio.h"
+#include "init.h"
+#include "user.h"
+#include "mode.h"
+
+struct aio_thread_req {
+        enum aio_type type;
+        int io_fd;
+        unsigned long long offset;
+        char *buf;
+        int len;
+        struct aio_context *aio;
+};
+
+static int aio_req_fd_r = -1;
+static int aio_req_fd_w = -1;
+
+#if defined(HAVE_AIO_ABI)
+#include <linux/aio_abi.h>
+
+/* If we have the headers, we are going to build with AIO enabled.
+ * If we don't have aio in libc, we define the necessary stubs here.
+ */
+
+#if !defined(HAVE_AIO_LIBC)
+
+static long io_setup(int n, aio_context_t *ctxp)
+{
+        return syscall(__NR_io_setup, n, ctxp);
+}
+
+static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
+{
+        return syscall(__NR_io_submit, ctx, nr, iocbpp);
+}
+
+static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
+                         struct io_event *events, struct timespec *timeout)
+{
+        return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
+}
+
+#endif
+
+/* The AIO_MMAP cases force the mmapped page into memory here
+ * rather than in whatever place first touches the data.  I used
+ * to do this by touching the page, but that's delicate because
+ * gcc is prone to optimizing that away.  So, what's done here
+ * is we read from the descriptor from which the page was
+ * mapped.  The caller is required to pass an offset which is
+ * inside the page that was mapped.  Thus, when the read
+ * returns, we know that the page is in the page cache, and
+ * that it now backs the mmapped area.
+ */
+
+static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
+                  int len, unsigned long long offset, struct aio_context *aio)
+{
+        struct iocb iocb, *iocbp = &iocb;
+        char c;
+        int err;
+
+        iocb = ((struct iocb) { .aio_data      = (unsigned long) aio,
+                                .aio_reqprio   = 0,
+                                .aio_fildes    = fd,
+                                .aio_buf       = (unsigned long) buf,
+                                .aio_nbytes    = len,
+                                .aio_offset    = offset,
+                                .aio_reserved1 = 0,
+                                .aio_reserved2 = 0,
+                                .aio_reserved3 = 0 });
+
+        switch(type){
+        case AIO_READ:
+                iocb.aio_lio_opcode = IOCB_CMD_PREAD;
+                err = io_submit(ctx, 1, &iocbp);
+                break;
+        case AIO_WRITE:
+                iocb.aio_lio_opcode = IOCB_CMD_PWRITE;
+                err = io_submit(ctx, 1, &iocbp);
+                break;
+        case AIO_MMAP:
+                iocb.aio_lio_opcode = IOCB_CMD_PREAD;
+                iocb.aio_buf = (unsigned long) &c;
+                iocb.aio_nbytes = sizeof(c);
+                err = io_submit(ctx, 1, &iocbp);
+                break;
+        default:
+                printk("Bogus op in do_aio - %d\n", type);
+                err = -EINVAL;
+                break;
+        }
+        if(err > 0)
+                err = 0;
+
+        return err;
+}
+
+static aio_context_t ctx = 0;
+
+static int aio_thread(void *arg)
+{
+        struct aio_thread_reply reply;
+        struct io_event event;
+        int err, n, reply_fd;
+
+        signal(SIGWINCH, SIG_IGN);
+
+        while(1){
+                n = io_getevents(ctx, 1, 1, &event, NULL);
+                if(n < 0){
+                        if(errno == EINTR)
+                                continue;
+                        printk("aio_thread - io_getevents failed, "
+                               "errno = %d\n", errno);
+                }
+                else {
+                        reply = ((struct aio_thread_reply)
+                                { .data = (void *) (long) event.data,
+                                  .err = event.res });
+                        reply_fd = ((struct aio_context *) reply.data)->reply_fd;
+                        err = os_write_file(reply_fd, &reply, sizeof(reply));
+                        if(err != sizeof(reply))
+                                printk("not_aio_thread - write failed, "
+                                       "fd = %d, err = %d\n",
+                                       aio_req_fd_r, -err);
+                }
+        }
+        return 0;
+}
+
+#endif
+
+static int do_not_aio(struct aio_thread_req *req)
+{
+        char c;
+        int err;
+
+        switch(req->type){
+        case AIO_READ:
+                err = os_seek_file(req->io_fd, req->offset);
+                if(err)
+                        goto out;
+
+                err = os_read_file(req->io_fd, req->buf, req->len);
+                break;
+        case AIO_WRITE:
+                err = os_seek_file(req->io_fd, req->offset);
+                if(err)
+                        goto out;
+
+                err = os_write_file(req->io_fd, req->buf, req->len);
+                break;
+        case AIO_MMAP:
+                err = os_seek_file(req->io_fd, req->offset);
+                if(err)
+                        goto out;
+
+                err = os_read_file(req->io_fd, &c, sizeof(c));
+                break;
+        default:
+                printk("do_not_aio - bad request type : %d\n", req->type);
+                err = -EINVAL;
+                break;
+        }
+
+ out:
+        return err;
+}
+
+static int not_aio_thread(void *arg)
+{
+        struct aio_thread_req req;
+        struct aio_thread_reply reply;
+        int err;
+
+        signal(SIGWINCH, SIG_IGN);
+        while(1){
+                err = os_read_file(aio_req_fd_r, &req, sizeof(req));
+                if(err != sizeof(req)){
+                        if(err < 0)
+                                printk("not_aio_thread - read failed, "
+                                       "fd = %d, err = %d\n", aio_req_fd_r,
+                                       -err);
+                        else {
+                                printk("not_aio_thread - short read, fd = %d, "
+                                       "length = %d\n", aio_req_fd_r, err);
+                        }
+                        continue;
+                }
+                err = do_not_aio(&req);
+                reply = ((struct aio_thread_reply) { .data     = req.aio,
+                                                     .err      = err });
+                err = os_write_file(req.aio->reply_fd, &reply, sizeof(reply));
+                if(err != sizeof(reply))
+                        printk("not_aio_thread - write failed, fd = %d, "
+                               "err = %d\n", aio_req_fd_r, -err);
+        }
+}
+
+static int aio_pid = -1;
+
+static int init_aio_24(void)
+{
+        unsigned long stack;
+        int fds[2], err;
+
+        err = os_pipe(fds, 1, 1);
+        if(err)
+                goto out;
+
+        aio_req_fd_w = fds[0];
+        aio_req_fd_r = fds[1];
+        err = run_helper_thread(not_aio_thread, NULL,
+                                CLONE_FILES | CLONE_VM | SIGCHLD, &stack, 0);
+        if(err < 0)
+                goto out_close_pipe;
+
+        aio_pid = err;
+        goto out;
+
+ out_close_pipe:
+        os_close_file(fds[0]);
+        os_close_file(fds[1]);
+        aio_req_fd_w = -1;
+        aio_req_fd_r = -1;
+ out:
+#ifndef HAVE_AIO_ABI
+       printk("/usr/include/linux/aio_abi.h not present during build\n");
+#endif
+       printk("2.6 host AIO support not used - falling back to I/O "
+              "thread\n");
+        return 0;
+}
+
+#ifdef HAVE_AIO_ABI
+#define DEFAULT_24_AIO 0
+static int init_aio_26(void)
+{
+        unsigned long stack;
+        int err;
+
+        if(io_setup(256, &ctx)){
+                printk("aio_thread failed to initialize context, err = %d\n",
+                       errno);
+                return -errno;
+        }
+
+        err = run_helper_thread(aio_thread, NULL,
+                                CLONE_FILES | CLONE_VM | SIGCHLD, &stack, 0);
+        if(err < 0)
+                return -errno;
+
+        aio_pid = err;
+
+       printk("Using 2.6 host AIO\n");
+        return 0;
+}
+
+static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
+                        unsigned long long offset, struct aio_context *aio)
+{
+        struct aio_thread_reply reply;
+        int err;
+
+        err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
+        if(err){
+                reply = ((struct aio_thread_reply) { .data = aio,
+                                                     .err  = err });
+                err = os_write_file(aio->reply_fd, &reply, sizeof(reply));
+                if(err != sizeof(reply))
+                        printk("submit_aio_26 - write failed, "
+                               "fd = %d, err = %d\n", aio->reply_fd, -err);
+                else err = 0;
+        }
+
+        return err;
+}
+
+#else
+#define DEFAULT_24_AIO 1
+static int init_aio_26(void)
+{
+        return -ENOSYS;
+}
+
+static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
+                        unsigned long long offset, struct aio_context *aio)
+{
+        return -ENOSYS;
+}
+#endif
+
+static int aio_24 = DEFAULT_24_AIO;
+
+static int __init set_aio_24(char *name, int *add)
+{
+        aio_24 = 1;
+        return 0;
+}
+
+__uml_setup("aio=2.4", set_aio_24,
+"aio=2.4\n"
+"    This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
+"    available.  2.4 AIO is a single thread that handles one request at a\n"
+"    time, synchronously.  2.6 AIO is a thread which uses the 2.6 AIO \n"
+"    interface to handle an arbitrary number of pending requests.  2.6 AIO \n"
+"    is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
+"    /usr/include/linux/aio_abi.h not available.  Many distributions don't\n"
+"    include aio_abi.h, so you will need to copy it from a kernel tree to\n"
+"    your /usr/include/linux in order to build an AIO-capable UML\n\n"
+);
+
+static int init_aio(void)
+{
+        int err;
+
+        CHOOSE_MODE(({
+                if(!aio_24){
+                        printk("Disabling 2.6 AIO in tt mode\n");
+                        aio_24 = 1;
+                } }), (void) 0);
+
+        if(!aio_24){
+                err = init_aio_26();
+                if(err && (errno == ENOSYS)){
+                        printk("2.6 AIO not supported on the host - "
+                               "reverting to 2.4 AIO\n");
+                        aio_24 = 1;
+                }
+                else return err;
+        }
+
+        if(aio_24)
+                return init_aio_24();
+
+        return 0;
+}
+
+/* The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
+ * needs to be called when the kernel is running because it calls run_helper,
+ * which needs get_free_page.  exit_aio is a __uml_exitcall because the generic
+ * kernel does not run __exitcalls on shutdown, and can't because many of them
+ * break when called outside of module unloading.
+ */
+__initcall(init_aio);
+
+static void exit_aio(void)
+{
+        if(aio_pid != -1)
+                os_kill_process(aio_pid, 1);
+}
+
+__uml_exitcall(exit_aio);
+
+static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
+                        unsigned long long offset, struct aio_context *aio)
+{
+        struct aio_thread_req req = { .type            = type,
+                                      .io_fd           = io_fd,
+                                      .offset          = offset,
+                                      .buf             = buf,
+                                      .len             = len,
+                                      .aio             = aio,
+        };
+        int err;
+
+        err = os_write_file(aio_req_fd_w, &req, sizeof(req));
+        if(err == sizeof(req))
+                err = 0;
+
+        return err;
+}
+
+int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
+               unsigned long long offset, int reply_fd,
+               struct aio_context *aio)
+{
+        aio->reply_fd = reply_fd;
+        if(aio_24)
+                return submit_aio_24(type, io_fd, buf, len, offset, aio);
+        else {
+                return submit_aio_26(type, io_fd, buf, len, offset, aio);
+        }
+}