From: Rusty Russell Date: Thu, 26 Jul 2007 17:41:03 +0000 (-0700) Subject: lguest: documentation IV: Launcher X-Git-Url: https://git.karo-electronics.de/?a=commitdiff_plain;h=dde797899ac17ebb812b7566044124d785e98dc7;p=linux-beck.git lguest: documentation IV: Launcher Documentation: The Launcher Signed-off-by: Rusty Russell Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- diff --git a/Documentation/lguest/lguest.c b/Documentation/lguest/lguest.c index fc1bf70abfb1..d7e26f025959 100644 --- a/Documentation/lguest/lguest.c +++ b/Documentation/lguest/lguest.c @@ -34,12 +34,20 @@ #include #include #include +/*L:110 We can ignore the 28 include files we need for this program, but I do + * want to draw attention to the use of kernel-style types. + * + * As Linus said, "C is a Spartan language, and so should your naming be." I + * like these abbreviations and the header we need uses them, so we define them + * here. + */ typedef unsigned long long u64; typedef uint32_t u32; typedef uint16_t u16; typedef uint8_t u8; #include "../../include/linux/lguest_launcher.h" #include "../../include/asm-i386/e820.h" +/*:*/ #define PAGE_PRESENT 0x7 /* Present, RW, Execute */ #define NET_PEERNUM 1 @@ -48,33 +56,52 @@ typedef uint8_t u8; #define SIOCBRADDIF 0x89a2 /* add interface to bridge */ #endif +/*L:120 verbose is both a global flag and a macro. The C preprocessor allows + * this, and although I wouldn't recommend it, it works quite nicely here. */ static bool verbose; #define verbose(args...) \ do { if (verbose) printf(args); } while(0) +/*:*/ + +/* The pipe to send commands to the waker process */ static int waker_fd; +/* The top of guest physical memory. */ static u32 top; +/* This is our list of devices. */ struct device_list { + /* Summary information about the devices in our list: ready to pass to + * select() to ask which need servicing.*/ fd_set infds; int max_infd; + /* The descriptor page for the devices. */ struct lguest_device_desc *descs; + + /* A single linked list of devices. */ struct device *dev; + /* ... And an end pointer so we can easily append new devices */ struct device **lastdev; }; +/* The device structure describes a single device. */ struct device { + /* The linked-list pointer. */ struct device *next; + /* The descriptor for this device, as mapped into the Guest. */ struct lguest_device_desc *desc; + /* The memory page(s) of this device, if any. Also mapped in Guest. */ void *mem; - /* Watch this fd if handle_input non-NULL. */ + /* If handle_input is set, it wants to be called when this file + * descriptor is ready. */ int fd; bool (*handle_input)(int fd, struct device *me); - /* Watch DMA to this key if handle_input non-NULL. */ + /* If handle_output is set, it wants to be called when the Guest sends + * DMA to this key. */ unsigned long watch_key; u32 (*handle_output)(int fd, const struct iovec *iov, unsigned int num, struct device *me); @@ -83,6 +110,11 @@ struct device void *priv; }; +/*L:130 + * Loading the Kernel. + * + * We start with couple of simple helper routines. open_or_die() avoids + * error-checking code cluttering the callers: */ static int open_or_die(const char *name, int flags) { int fd = open(name, flags); @@ -91,26 +123,38 @@ static int open_or_die(const char *name, int flags) return fd; } +/* map_zeroed_pages() takes a (page-aligned) address and a number of pages. */ static void *map_zeroed_pages(unsigned long addr, unsigned int num) { + /* We cache the /dev/zero file-descriptor so we only open it once. */ static int fd = -1; if (fd == -1) fd = open_or_die("/dev/zero", O_RDONLY); + /* We use a private mapping (ie. if we write to the page, it will be + * copied), and obviously we insist that it be mapped where we ask. */ if (mmap((void *)addr, getpagesize() * num, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_FIXED|MAP_PRIVATE, fd, 0) != (void *)addr) err(1, "Mmaping %u pages of /dev/zero @%p", num, (void *)addr); + + /* Returning the address is just a courtesy: can simplify callers. */ return (void *)addr; } -/* Find magic string marking entry point, return entry point. */ +/* To find out where to start we look for the magic Guest string, which marks + * the code we see in lguest_asm.S. This is a hack which we are currently + * plotting to replace with the normal Linux entry point. */ static unsigned long entry_point(void *start, void *end, unsigned long page_offset) { void *p; + /* The scan gives us the physical starting address. We want the + * virtual address in this case, and fortunately, we already figured + * out the physical-virtual difference and passed it here in + * "page_offset". */ for (p = start; p < end; p++) if (memcmp(p, "GenuineLguest", strlen("GenuineLguest")) == 0) return (long)p + strlen("GenuineLguest") + page_offset; @@ -118,7 +162,17 @@ static unsigned long entry_point(void *start, void *end, err(1, "Is this image a genuine lguest?"); } -/* Returns the entry point */ +/* This routine takes an open vmlinux image, which is in ELF, and maps it into + * the Guest memory. ELF = Embedded Linking Format, which is the format used + * by all modern binaries on Linux including the kernel. + * + * The ELF headers give *two* addresses: a physical address, and a virtual + * address. The Guest kernel expects to be placed in memory at the physical + * address, and the page tables set up so it will correspond to that virtual + * address. We return the difference between the virtual and physical + * addresses in the "page_offset" pointer. + * + * We return the starting address. */ static unsigned long map_elf(int elf_fd, const Elf32_Ehdr *ehdr, unsigned long *page_offset) { @@ -127,40 +181,61 @@ static unsigned long map_elf(int elf_fd, const Elf32_Ehdr *ehdr, unsigned int i; unsigned long start = -1UL, end = 0; - /* Sanity checks. */ + /* Sanity checks on the main ELF header: an x86 executable with a + * reasonable number of correctly-sized program headers. */ if (ehdr->e_type != ET_EXEC || ehdr->e_machine != EM_386 || ehdr->e_phentsize != sizeof(Elf32_Phdr) || ehdr->e_phnum < 1 || ehdr->e_phnum > 65536U/sizeof(Elf32_Phdr)) errx(1, "Malformed elf header"); + /* An ELF executable contains an ELF header and a number of "program" + * headers which indicate which parts ("segments") of the program to + * load where. */ + + /* We read in all the program headers at once: */ if (lseek(elf_fd, ehdr->e_phoff, SEEK_SET) < 0) err(1, "Seeking to program headers"); if (read(elf_fd, phdr, sizeof(phdr)) != sizeof(phdr)) err(1, "Reading program headers"); + /* We don't know page_offset yet. */ *page_offset = 0; - /* We map the loadable segments at virtual addresses corresponding - * to their physical addresses (our virtual == guest physical). */ + + /* Try all the headers: there are usually only three. A read-only one, + * a read-write one, and a "note" section which isn't loadable. */ for (i = 0; i < ehdr->e_phnum; i++) { + /* If this isn't a loadable segment, we ignore it */ if (phdr[i].p_type != PT_LOAD) continue; verbose("Section %i: size %i addr %p\n", i, phdr[i].p_memsz, (void *)phdr[i].p_paddr); - /* We expect linear address space. */ + /* We expect a simple linear address space: every segment must + * have the same difference between virtual (p_vaddr) and + * physical (p_paddr) address. */ if (!*page_offset) *page_offset = phdr[i].p_vaddr - phdr[i].p_paddr; else if (*page_offset != phdr[i].p_vaddr - phdr[i].p_paddr) errx(1, "Page offset of section %i different", i); + /* We track the first and last address we mapped, so we can + * tell entry_point() where to scan. */ if (phdr[i].p_paddr < start) start = phdr[i].p_paddr; if (phdr[i].p_paddr + phdr[i].p_filesz > end) end = phdr[i].p_paddr + phdr[i].p_filesz; - /* We map everything private, writable. */ + /* We map this section of the file at its physical address. We + * map it read & write even if the header says this segment is + * read-only. The kernel really wants to be writable: it + * patches its own instructions which would normally be + * read-only. + * + * MAP_PRIVATE means that the page won't be copied until a + * write is done to it. This allows us to share much of the + * kernel memory between Guests. */ addr = mmap((void *)phdr[i].p_paddr, phdr[i].p_filesz, PROT_READ|PROT_WRITE|PROT_EXEC, @@ -174,7 +249,31 @@ static unsigned long map_elf(int elf_fd, const Elf32_Ehdr *ehdr, return entry_point((void *)start, (void *)end, *page_offset); } -/* This is amazingly reliable. */ +/*L:170 Prepare to be SHOCKED and AMAZED. And possibly a trifle nauseated. + * + * We know that CONFIG_PAGE_OFFSET sets what virtual address the kernel expects + * to be. We don't know what that option was, but we can figure it out + * approximately by looking at the addresses in the code. I chose the common + * case of reading a memory location into the %eax register: + * + * movl , %eax + * + * This gets encoded as five bytes: "0xA1 <4-byte-address>". For example, + * "0xA1 0x18 0x60 0x47 0xC0" reads the address 0xC0476018 into %eax. + * + * In this example can guess that the kernel was compiled with + * CONFIG_PAGE_OFFSET set to 0xC0000000 (it's always a round number). If the + * kernel were larger than 16MB, we might see 0xC1 addresses show up, but our + * kernel isn't that bloated yet. + * + * Unfortunately, x86 has variable-length instructions, so finding this + * particular instruction properly involves writing a disassembler. Instead, + * we rely on statistics. We look for "0xA1" and tally the different bytes + * which occur 4 bytes later (the "0xC0" in our example above). When one of + * those bytes appears three times, we can be reasonably confident that it + * forms the start of CONFIG_PAGE_OFFSET. + * + * This is amazingly reliable. */ static unsigned long intuit_page_offset(unsigned char *img, unsigned long len) { unsigned int i, possibilities[256] = { 0 }; @@ -187,30 +286,52 @@ static unsigned long intuit_page_offset(unsigned char *img, unsigned long len) errx(1, "could not determine page offset"); } +/*L:160 Unfortunately the entire ELF image isn't compressed: the segments + * which need loading are extracted and compressed raw. This denies us the + * information we need to make a fully-general loader. */ static unsigned long unpack_bzimage(int fd, unsigned long *page_offset) { gzFile f; int ret, len = 0; + /* A bzImage always gets loaded at physical address 1M. This is + * actually configurable as CONFIG_PHYSICAL_START, but as the comment + * there says, "Don't change this unless you know what you are doing". + * Indeed. */ void *img = (void *)0x100000; + /* gzdopen takes our file descriptor (carefully placed at the start of + * the GZIP header we found) and returns a gzFile. */ f = gzdopen(fd, "rb"); + /* We read it into memory in 64k chunks until we hit the end. */ while ((ret = gzread(f, img + len, 65536)) > 0) len += ret; if (ret < 0) err(1, "reading image from bzImage"); verbose("Unpacked size %i addr %p\n", len, img); + + /* Without the ELF header, we can't tell virtual-physical gap. This is + * CONFIG_PAGE_OFFSET, and people do actually change it. Fortunately, + * I have a clever way of figuring it out from the code itself. */ *page_offset = intuit_page_offset(img, len); return entry_point(img, img + len, *page_offset); } +/*L:150 A bzImage, unlike an ELF file, is not meant to be loaded. You're + * supposed to jump into it and it will unpack itself. We can't do that + * because the Guest can't run the unpacking code, and adding features to + * lguest kills puppies, so we don't want to. + * + * The bzImage is formed by putting the decompressing code in front of the + * compressed kernel code. So we can simple scan through it looking for the + * first "gzip" header, and start decompressing from there. */ static unsigned long load_bzimage(int fd, unsigned long *page_offset) { unsigned char c; int state = 0; - /* Ugly brute force search for gzip header. */ + /* GZIP header is 0x1F 0x8B ... . */ while (read(fd, &c, 1) == 1) { switch (state) { case 0: @@ -227,8 +348,10 @@ static unsigned long load_bzimage(int fd, unsigned long *page_offset) state++; break; case 9: + /* Seek back to the start of the gzip header. */ lseek(fd, -10, SEEK_CUR); - if (c != 0x03) /* Compressed under UNIX. */ + /* One final check: "compressed under UNIX". */ + if (c != 0x03) state = -1; else return unpack_bzimage(fd, page_offset); @@ -237,25 +360,43 @@ static unsigned long load_bzimage(int fd, unsigned long *page_offset) errx(1, "Could not find kernel in bzImage"); } +/*L:140 Loading the kernel is easy when it's a "vmlinux", but most kernels + * come wrapped up in the self-decompressing "bzImage" format. With some funky + * coding, we can load those, too. */ static unsigned long load_kernel(int fd, unsigned long *page_offset) { Elf32_Ehdr hdr; + /* Read in the first few bytes. */ if (read(fd, &hdr, sizeof(hdr)) != sizeof(hdr)) err(1, "Reading kernel"); + /* If it's an ELF file, it starts with "\177ELF" */ if (memcmp(hdr.e_ident, ELFMAG, SELFMAG) == 0) return map_elf(fd, &hdr, page_offset); + /* Otherwise we assume it's a bzImage, and try to unpack it */ return load_bzimage(fd, page_offset); } +/* This is a trivial little helper to align pages. Andi Kleen hated it because + * it calls getpagesize() twice: "it's dumb code." + * + * Kernel guys get really het up about optimization, even when it's not + * necessary. I leave this code as a reaction against that. */ static inline unsigned long page_align(unsigned long addr) { + /* Add upwards and truncate downwards. */ return ((addr + getpagesize()-1) & ~(getpagesize()-1)); } -/* initrd gets loaded at top of memory: return length. */ +/*L:180 An "initial ram disk" is a disk image loaded into memory along with + * the kernel which the kernel can use to boot from without needing any + * drivers. Most distributions now use this as standard: the initrd contains + * the code to load the appropriate driver modules for the current machine. + * + * Importantly, James Morris works for RedHat, and Fedora uses initrds for its + * kernels. He sent me this (and tells me when I break it). */ static unsigned long load_initrd(const char *name, unsigned long mem) { int ifd; @@ -264,21 +405,35 @@ static unsigned long load_initrd(const char *name, unsigned long mem) void *iaddr; ifd = open_or_die(name, O_RDONLY); + /* fstat() is needed to get the file size. */ if (fstat(ifd, &st) < 0) err(1, "fstat() on initrd '%s'", name); + /* The length needs to be rounded up to a page size: mmap needs the + * address to be page aligned. */ len = page_align(st.st_size); + /* We map the initrd at the top of memory. */ iaddr = mmap((void *)mem - len, st.st_size, PROT_READ|PROT_EXEC|PROT_WRITE, MAP_FIXED|MAP_PRIVATE, ifd, 0); if (iaddr != (void *)mem - len) err(1, "Mmaping initrd '%s' returned %p not %p", name, iaddr, (void *)mem - len); + /* Once a file is mapped, you can close the file descriptor. It's a + * little odd, but quite useful. */ close(ifd); verbose("mapped initrd %s size=%lu @ %p\n", name, st.st_size, iaddr); + + /* We return the initrd size. */ return len; } +/* Once we know how much memory we have, and the address the Guest kernel + * expects, we can construct simple linear page tables which will get the Guest + * far enough into the boot to create its own. + * + * We lay them out of the way, just below the initrd (which is why we need to + * know its size). */ static unsigned long setup_pagetables(unsigned long mem, unsigned long initrd_size, unsigned long page_offset) @@ -287,23 +442,32 @@ static unsigned long setup_pagetables(unsigned long mem, unsigned int mapped_pages, i, linear_pages; unsigned int ptes_per_page = getpagesize()/sizeof(u32); - /* If we can map all of memory above page_offset, we do so. */ + /* Ideally we map all physical memory starting at page_offset. + * However, if page_offset is 0xC0000000 we can only map 1G of physical + * (0xC0000000 + 1G overflows). */ if (mem <= -page_offset) mapped_pages = mem/getpagesize(); else mapped_pages = -page_offset/getpagesize(); - /* Each linear PTE page can map ptes_per_page pages. */ + /* Each PTE page can map ptes_per_page pages: how many do we need? */ linear_pages = (mapped_pages + ptes_per_page-1)/ptes_per_page; - /* We lay out top-level then linear mapping immediately below initrd */ + /* We put the toplevel page directory page at the top of memory. */ pgdir = (void *)mem - initrd_size - getpagesize(); + + /* Now we use the next linear_pages pages as pte pages */ linear = (void *)pgdir - linear_pages*getpagesize(); + /* Linear mapping is easy: put every page's address into the mapping in + * order. PAGE_PRESENT contains the flags Present, Writable and + * Executable. */ for (i = 0; i < mapped_pages; i++) linear[i] = ((i * getpagesize()) | PAGE_PRESENT); - /* Now set up pgd so that this memory is at page_offset */ + /* The top level points to the linear page table pages above. The + * entry representing page_offset points to the first one, and they + * continue from there. */ for (i = 0; i < mapped_pages; i += ptes_per_page) { pgdir[(i + page_offset/getpagesize())/ptes_per_page] = (((u32)linear + i*sizeof(u32)) | PAGE_PRESENT); @@ -312,9 +476,13 @@ static unsigned long setup_pagetables(unsigned long mem, verbose("Linear mapping of %u pages in %u pte pages at %p\n", mapped_pages, linear_pages, linear); + /* We return the top level (guest-physical) address: the kernel needs + * to know where it is. */ return (unsigned long)pgdir; } +/* Simple routine to roll all the commandline arguments together with spaces + * between them. */ static void concat(char *dst, char *args[]) { unsigned int i, len = 0; @@ -328,6 +496,10 @@ static void concat(char *dst, char *args[]) dst[len] = '\0'; } +/* This is where we actually tell the kernel to initialize the Guest. We saw + * the arguments it expects when we looked at initialize() in lguest_user.c: + * the top physical page to allow, the top level pagetable, the entry point and + * the page_offset constant for the Guest. */ static int tell_kernel(u32 pgdir, u32 start, u32 page_offset) { u32 args[] = { LHREQ_INITIALIZE, @@ -337,8 +509,11 @@ static int tell_kernel(u32 pgdir, u32 start, u32 page_offset) fd = open_or_die("/dev/lguest", O_RDWR); if (write(fd, args, sizeof(args)) < 0) err(1, "Writing to /dev/lguest"); + + /* We return the /dev/lguest file descriptor to control this Guest */ return fd; } +/*:*/ static void set_fd(int fd, struct device_list *devices) { @@ -347,61 +522,108 @@ static void set_fd(int fd, struct device_list *devices) devices->max_infd = fd; } -/* When input arrives, we tell the kernel to kick lguest out with -EAGAIN. */ +/*L:200 + * The Waker. + * + * With a console and network devices, we can have lots of input which we need + * to process. We could try to tell the kernel what file descriptors to watch, + * but handing a file descriptor mask through to the kernel is fairly icky. + * + * Instead, we fork off a process which watches the file descriptors and writes + * the LHREQ_BREAK command to the /dev/lguest filedescriptor to tell the Host + * loop to stop running the Guest. This causes it to return from the + * /dev/lguest read with -EAGAIN, where it will write to /dev/lguest to reset + * the LHREQ_BREAK and wake us up again. + * + * This, of course, is merely a different *kind* of icky. + */ static void wake_parent(int pipefd, int lguest_fd, struct device_list *devices) { + /* Add the pipe from the Launcher to the fdset in the device_list, so + * we watch it, too. */ set_fd(pipefd, devices); for (;;) { fd_set rfds = devices->infds; u32 args[] = { LHREQ_BREAK, 1 }; + /* Wait until input is ready from one of the devices. */ select(devices->max_infd+1, &rfds, NULL, NULL, NULL); + /* Is it a message from the Launcher? */ if (FD_ISSET(pipefd, &rfds)) { int ignorefd; + /* If read() returns 0, it means the Launcher has + * exited. We silently follow. */ if (read(pipefd, &ignorefd, sizeof(ignorefd)) == 0) exit(0); + /* Otherwise it's telling us there's a problem with one + * of the devices, and we should ignore that file + * descriptor from now on. */ FD_CLR(ignorefd, &devices->infds); - } else + } else /* Send LHREQ_BREAK command. */ write(lguest_fd, args, sizeof(args)); } } +/* This routine just sets up a pipe to the Waker process. */ static int setup_waker(int lguest_fd, struct device_list *device_list) { int pipefd[2], child; + /* We create a pipe to talk to the waker, and also so it knows when the + * Launcher dies (and closes pipe). */ pipe(pipefd); child = fork(); if (child == -1) err(1, "forking"); if (child == 0) { + /* Close the "writing" end of our copy of the pipe */ close(pipefd[1]); wake_parent(pipefd[0], lguest_fd, device_list); } + /* Close the reading end of our copy of the pipe. */ close(pipefd[0]); + /* Here is the fd used to talk to the waker. */ return pipefd[1]; } +/*L:210 + * Device Handling. + * + * When the Guest sends DMA to us, it sends us an array of addresses and sizes. + * We need to make sure it's not trying to reach into the Launcher itself, so + * we have a convenient routine which check it and exits with an error message + * if something funny is going on: + */ static void *_check_pointer(unsigned long addr, unsigned int size, unsigned int line) { + /* We have to separately check addr and addr+size, because size could + * be huge and addr + size might wrap around. */ if (addr >= top || addr + size >= top) errx(1, "%s:%i: Invalid address %li", __FILE__, line, addr); + /* We return a pointer for the caller's convenience, now we know it's + * safe to use. */ return (void *)addr; } +/* A macro which transparently hands the line number to the real function. */ #define check_pointer(addr,size) _check_pointer(addr, size, __LINE__) -/* Returns pointer to dma->used_len */ +/* The Guest has given us the address of a "struct lguest_dma". We check it's + * OK and convert it to an iovec (which is a simple array of ptr/size + * pairs). */ static u32 *dma2iov(unsigned long dma, struct iovec iov[], unsigned *num) { unsigned int i; struct lguest_dma *udma; + /* First we make sure that the array memory itself is valid. */ udma = check_pointer(dma, sizeof(*udma)); + /* Now we check each element */ for (i = 0; i < LGUEST_MAX_DMA_SECTIONS; i++) { + /* A zero length ends the array. */ if (!udma->len[i]) break; @@ -409,9 +631,15 @@ static u32 *dma2iov(unsigned long dma, struct iovec iov[], unsigned *num) iov[i].iov_len = udma->len[i]; } *num = i; + + /* We return the pointer to where the caller should write the amount of + * the buffer used. */ return &udma->used_len; } +/* This routine gets a DMA buffer from the Guest for a given key, and converts + * it to an iovec array. It returns the interrupt the Guest wants when we're + * finished, and a pointer to the "used_len" field to fill in. */ static u32 *get_dma_buffer(int fd, void *key, struct iovec iov[], unsigned int *num, u32 *irq) { @@ -419,16 +647,21 @@ static u32 *get_dma_buffer(int fd, void *key, unsigned long udma; u32 *res; + /* Ask the kernel for a DMA buffer corresponding to this key. */ udma = write(fd, buf, sizeof(buf)); + /* They haven't registered any, or they're all used? */ if (udma == (unsigned long)-1) return NULL; - /* Kernel stashes irq in ->used_len. */ + /* Convert it into our iovec array */ res = dma2iov(udma, iov, num); + /* The kernel stashes irq in ->used_len to get it out to us. */ *irq = *res; + /* Return a pointer to ((struct lguest_dma *)udma)->used_len. */ return res; } +/* This is a convenient routine to send the Guest an interrupt. */ static void trigger_irq(int fd, u32 irq) { u32 buf[] = { LHREQ_IRQ, irq }; @@ -436,6 +669,10 @@ static void trigger_irq(int fd, u32 irq) err(1, "Triggering irq %i", irq); } +/* This simply sets up an iovec array where we can put data to be discarded. + * This happens when the Guest doesn't want or can't handle the input: we have + * to get rid of it somewhere, and if we bury it in the ceiling space it will + * start to smell after a week. */ static void discard_iovec(struct iovec *iov, unsigned int *num) { static char discard_buf[1024]; @@ -444,19 +681,24 @@ static void discard_iovec(struct iovec *iov, unsigned int *num) iov->iov_len = sizeof(discard_buf); } +/* Here is the input terminal setting we save, and the routine to restore them + * on exit so the user can see what they type next. */ static struct termios orig_term; static void restore_term(void) { tcsetattr(STDIN_FILENO, TCSANOW, &orig_term); } +/* We associate some data with the console for our exit hack. */ struct console_abort { + /* How many times have they hit ^C? */ int count; + /* When did they start? */ struct timeval start; }; -/* We DMA input to buffer bound at start of console page. */ +/* This is the routine which handles console input (ie. stdin). */ static bool handle_console_input(int fd, struct device *dev) { u32 irq = 0, *lenp; @@ -465,24 +707,38 @@ static bool handle_console_input(int fd, struct device *dev) struct iovec iov[LGUEST_MAX_DMA_SECTIONS]; struct console_abort *abort = dev->priv; + /* First we get the console buffer from the Guest. The key is dev->mem + * which was set to 0 in setup_console(). */ lenp = get_dma_buffer(fd, dev->mem, iov, &num, &irq); if (!lenp) { + /* If it's not ready for input, warn and set up to discard. */ warn("console: no dma buffer!"); discard_iovec(iov, &num); } + /* This is why we convert to iovecs: the readv() call uses them, and so + * it reads straight into the Guest's buffer. */ len = readv(dev->fd, iov, num); if (len <= 0) { + /* This implies that the console is closed, is /dev/null, or + * something went terribly wrong. We still go through the rest + * of the logic, though, especially the exit handling below. */ warnx("Failed to get console input, ignoring console."); len = 0; } + /* If we read the data into the Guest, fill in the length and send the + * interrupt. */ if (lenp) { *lenp = len; trigger_irq(fd, irq); } - /* Three ^C within one second? Exit. */ + /* Three ^C within one second? Exit. + * + * This is such a hack, but works surprisingly well. Each ^C has to be + * in a buffer by itself, so they can't be too fast. But we check that + * we get three within about a second, so they can't be too slow. */ if (len == 1 && ((char *)iov[0].iov_base)[0] == 3) { if (!abort->count++) gettimeofday(&abort->start, NULL); @@ -490,43 +746,60 @@ static bool handle_console_input(int fd, struct device *dev) struct timeval now; gettimeofday(&now, NULL); if (now.tv_sec <= abort->start.tv_sec+1) { - /* Make sure waker is not blocked in BREAK */ u32 args[] = { LHREQ_BREAK, 0 }; + /* Close the fd so Waker will know it has to + * exit. */ close(waker_fd); + /* Just in case waker is blocked in BREAK, send + * unbreak now. */ write(fd, args, sizeof(args)); exit(2); } abort->count = 0; } } else + /* Any other key resets the abort counter. */ abort->count = 0; + /* Now, if we didn't read anything, put the input terminal back and + * return failure (meaning, don't call us again). */ if (!len) { restore_term(); return false; } + /* Everything went OK! */ return true; } +/* Handling console output is much simpler than input. */ static u32 handle_console_output(int fd, const struct iovec *iov, unsigned num, struct device*dev) { + /* Whatever the Guest sends, write it to standard output. Return the + * number of bytes written. */ return writev(STDOUT_FILENO, iov, num); } +/* Guest->Host network output is also pretty easy. */ static u32 handle_tun_output(int fd, const struct iovec *iov, unsigned num, struct device *dev) { - /* Now we've seen output, we should warn if we can't get buffers. */ + /* We put a flag in the "priv" pointer of the network device, and set + * it as soon as we see output. We'll see why in handle_tun_input() */ *(bool *)dev->priv = true; + /* Whatever packet the Guest sent us, write it out to the tun + * device. */ return writev(dev->fd, iov, num); } +/* This matches the peer_key() in lguest_net.c. The key for any given slot + * is the address of the network device's page plus 4 * the slot number. */ static unsigned long peer_offset(unsigned int peernum) { return 4 * peernum; } +/* This is where we handle a packet coming in from the tun device */ static bool handle_tun_input(int fd, struct device *dev) { u32 irq = 0, *lenp; @@ -534,17 +807,28 @@ static bool handle_tun_input(int fd, struct device *dev) unsigned num; struct iovec iov[LGUEST_MAX_DMA_SECTIONS]; + /* First we get a buffer the Guest has bound to its key. */ lenp = get_dma_buffer(fd, dev->mem+peer_offset(NET_PEERNUM), iov, &num, &irq); if (!lenp) { + /* Now, it's expected that if we try to send a packet too + * early, the Guest won't be ready yet. This is why we set a + * flag when the Guest sends its first packet. If it's sent a + * packet we assume it should be ready to receive them. + * + * Actually, this is what the status bits in the descriptor are + * for: we should *use* them. FIXME! */ if (*(bool *)dev->priv) warn("network: no dma buffer!"); discard_iovec(iov, &num); } + /* Read the packet from the device directly into the Guest's buffer. */ len = readv(dev->fd, iov, num); if (len <= 0) err(1, "reading network"); + + /* Write the used_len, and trigger the interrupt for the Guest */ if (lenp) { *lenp = len; trigger_irq(fd, irq); @@ -552,9 +836,13 @@ static bool handle_tun_input(int fd, struct device *dev) verbose("tun input packet len %i [%02x %02x] (%s)\n", len, ((u8 *)iov[0].iov_base)[0], ((u8 *)iov[0].iov_base)[1], lenp ? "sent" : "discarded"); + /* All good. */ return true; } +/* The last device handling routine is block output: the Guest has sent a DMA + * to the block device. It will have placed the command it wants in the + * "struct lguest_block_page". */ static u32 handle_block_output(int fd, const struct iovec *iov, unsigned num, struct device *dev) { @@ -564,36 +852,64 @@ static u32 handle_block_output(int fd, const struct iovec *iov, struct iovec reply[LGUEST_MAX_DMA_SECTIONS]; off64_t device_len, off = (off64_t)p->sector * 512; + /* First we extract the device length from the dev->priv pointer. */ device_len = *(off64_t *)dev->priv; + /* We first check that the read or write is within the length of the + * block file. */ if (off >= device_len) err(1, "Bad offset %llu vs %llu", off, device_len); + /* Move to the right location in the block file. This shouldn't fail, + * but best to check. */ if (lseek64(dev->fd, off, SEEK_SET) != off) err(1, "Bad seek to sector %i", p->sector); verbose("Block: %s at offset %llu\n", p->type ? "WRITE" : "READ", off); + /* They were supposed to bind a reply buffer at key equal to the start + * of the block device memory. We need this to tell them when the + * request is finished. */ lenp = get_dma_buffer(fd, dev->mem, reply, &reply_num, &irq); if (!lenp) err(1, "Block request didn't give us a dma buffer"); if (p->type) { + /* A write request. The DMA they sent contained the data, so + * write it out. */ len = writev(dev->fd, iov, num); + /* Grr... Now we know how long the "struct lguest_dma" they + * sent was, we make sure they didn't try to write over the end + * of the block file (possibly extending it). */ if (off + len > device_len) { + /* Trim it back to the correct length */ ftruncate(dev->fd, device_len); + /* Die, bad Guest, die. */ errx(1, "Write past end %llu+%u", off, len); } + /* The reply length is 0: we just send back an empty DMA to + * interrupt them and tell them the write is finished. */ *lenp = 0; } else { + /* A read request. They sent an empty DMA to start the + * request, and we put the read contents into the reply + * buffer. */ len = readv(dev->fd, reply, reply_num); *lenp = len; } + /* The result is 1 (done), 2 if there was an error (short read or + * write). */ p->result = 1 + (p->bytes != len); + /* Now tell them we've used their reply buffer. */ trigger_irq(fd, irq); + + /* We're supposed to return the number of bytes of the output buffer we + * used. But the block device uses the "result" field instead, so we + * don't bother. */ return 0; } +/* This is the generic routine we call when the Guest sends some DMA out. */ static void handle_output(int fd, unsigned long dma, unsigned long key, struct device_list *devices) { @@ -602,30 +918,53 @@ static void handle_output(int fd, unsigned long dma, unsigned long key, struct iovec iov[LGUEST_MAX_DMA_SECTIONS]; unsigned num = 0; + /* Convert the "struct lguest_dma" they're sending to a "struct + * iovec". */ lenp = dma2iov(dma, iov, &num); + + /* Check each device: if they expect output to this key, tell them to + * handle it. */ for (i = devices->dev; i; i = i->next) { if (i->handle_output && key == i->watch_key) { + /* We write the result straight into the used_len field + * for them. */ *lenp = i->handle_output(fd, iov, num, i); return; } } + + /* This can happen: the kernel sends any SEND_DMA which doesn't match + * another Guest to us. It could be that another Guest just left a + * network, for example. But it's unusual. */ warnx("Pending dma %p, key %p", (void *)dma, (void *)key); } +/* This is called when the waker wakes us up: check for incoming file + * descriptors. */ static void handle_input(int fd, struct device_list *devices) { + /* select() wants a zeroed timeval to mean "don't wait". */ struct timeval poll = { .tv_sec = 0, .tv_usec = 0 }; for (;;) { struct device *i; fd_set fds = devices->infds; + /* If nothing is ready, we're done. */ if (select(devices->max_infd+1, &fds, NULL, NULL, &poll) == 0) break; + /* Otherwise, call the device(s) which have readable + * file descriptors and a method of handling them. */ for (i = devices->dev; i; i = i->next) { if (i->handle_input && FD_ISSET(i->fd, &fds)) { + /* If handle_input() returns false, it means we + * should no longer service it. + * handle_console_input() does this. */ if (!i->handle_input(fd, i)) { + /* Clear it from the set of input file + * descriptors kept at the head of the + * device list. */ FD_CLR(i->fd, &devices->infds); /* Tell waker to ignore it too... */ write(waker_fd, &i->fd, sizeof(i->fd)); @@ -635,6 +974,15 @@ static void handle_input(int fd, struct device_list *devices) } } +/*L:190 + * Device Setup + * + * All devices need a descriptor so the Guest knows it exists, and a "struct + * device" so the Launcher can keep track of it. We have common helper + * routines to allocate them. + * + * This routine allocates a new "struct lguest_device_desc" from descriptor + * table in the devices array just above the Guest's normal memory. */ static struct lguest_device_desc * new_dev_desc(struct lguest_device_desc *descs, u16 type, u16 features, u16 num_pages) @@ -646,6 +994,8 @@ new_dev_desc(struct lguest_device_desc *descs, descs[i].type = type; descs[i].features = features; descs[i].num_pages = num_pages; + /* If they said the device needs memory, we allocate + * that now, bumping up the top of Guest memory. */ if (num_pages) { map_zeroed_pages(top, num_pages); descs[i].pfn = top/getpagesize(); @@ -657,6 +1007,9 @@ new_dev_desc(struct lguest_device_desc *descs, errx(1, "too many devices"); } +/* This monster routine does all the creation and setup of a new device, + * including caling new_dev_desc() to allocate the descriptor and device + * memory. */ static struct device *new_device(struct device_list *devices, u16 type, u16 num_pages, u16 features, int fd, @@ -669,12 +1022,18 @@ static struct device *new_device(struct device_list *devices, { struct device *dev = malloc(sizeof(*dev)); - /* Append to device list. */ + /* Append to device list. Prepending to a single-linked list is + * easier, but the user expects the devices to be arranged on the bus + * in command-line order. The first network device on the command line + * is eth0, the first block device /dev/lgba, etc. */ *devices->lastdev = dev; dev->next = NULL; devices->lastdev = &dev->next; + /* Now we populate the fields one at a time. */ dev->fd = fd; + /* If we have an input handler for this file descriptor, then we add it + * to the device_list's fdset and maxfd. */ if (handle_input) set_fd(dev->fd, devices); dev->desc = new_dev_desc(devices->descs, type, features, num_pages); @@ -685,27 +1044,37 @@ static struct device *new_device(struct device_list *devices, return dev; } +/* Our first setup routine is the console. It's a fairly simple device, but + * UNIX tty handling makes it uglier than it could be. */ static void setup_console(struct device_list *devices) { struct device *dev; + /* If we can save the initial standard input settings... */ if (tcgetattr(STDIN_FILENO, &orig_term) == 0) { struct termios term = orig_term; + /* Then we turn off echo, line buffering and ^C etc. We want a + * raw input stream to the Guest. */ term.c_lflag &= ~(ISIG|ICANON|ECHO); tcsetattr(STDIN_FILENO, TCSANOW, &term); + /* If we exit gracefully, the original settings will be + * restored so the user can see what they're typing. */ atexit(restore_term); } - /* We don't currently require a page for the console. */ + /* We don't currently require any memory for the console, so we ask for + * 0 pages. */ dev = new_device(devices, LGUEST_DEVICE_T_CONSOLE, 0, 0, STDIN_FILENO, handle_console_input, LGUEST_CONSOLE_DMA_KEY, handle_console_output); + /* We store the console state in dev->priv, and initialize it. */ dev->priv = malloc(sizeof(struct console_abort)); ((struct console_abort *)dev->priv)->count = 0; verbose("device %p: console\n", (void *)(dev->desc->pfn * getpagesize())); } +/* Setting up a block file is also fairly straightforward. */ static void setup_block_file(const char *filename, struct device_list *devices) { int fd; @@ -713,20 +1082,47 @@ static void setup_block_file(const char *filename, struct device_list *devices) off64_t *device_len; struct lguest_block_page *p; + /* We open with O_LARGEFILE because otherwise we get stuck at 2G. We + * open with O_DIRECT because otherwise our benchmarks go much too + * fast. */ fd = open_or_die(filename, O_RDWR|O_LARGEFILE|O_DIRECT); + + /* We want one page, and have no input handler (the block file never + * has anything interesting to say to us). Our timing will be quite + * random, so it should be a reasonable randomness source. */ dev = new_device(devices, LGUEST_DEVICE_T_BLOCK, 1, LGUEST_DEVICE_F_RANDOMNESS, fd, NULL, 0, handle_block_output); + + /* We store the device size in the private area */ device_len = dev->priv = malloc(sizeof(*device_len)); + /* This is the safe way of establishing the size of our device: it + * might be a normal file or an actual block device like /dev/hdb. */ *device_len = lseek64(fd, 0, SEEK_END); - p = dev->mem; + /* The device memory is a "struct lguest_block_page". It's zeroed + * already, we just need to put in the device size. Block devices + * think in sectors (ie. 512 byte chunks), so we translate here. */ + p = dev->mem; p->num_sectors = *device_len/512; verbose("device %p: block %i sectors\n", (void *)(dev->desc->pfn * getpagesize()), p->num_sectors); } -/* We use fnctl locks to reserve network slots (autocleanup!) */ +/* + * Network Devices. + * + * Setting up network devices is quite a pain, because we have three types. + * First, we have the inter-Guest network. This is a file which is mapped into + * the address space of the Guests who are on the network. Because it is a + * shared mapping, the same page underlies all the devices, and they can send + * DMA to each other. + * + * Remember from our network driver, the Guest is told what slot in the page it + * is to use. We use exclusive fnctl locks to reserve a slot. If another + * Guest is using a slot, the lock will fail and we try another. Because fnctl + * locks are cleaned up automatically when we die, this cleverly means that our + * reservation on the slot will vanish if we crash. */ static unsigned int find_slot(int netfd, const char *filename) { struct flock fl; @@ -734,26 +1130,33 @@ static unsigned int find_slot(int netfd, const char *filename) fl.l_type = F_WRLCK; fl.l_whence = SEEK_SET; fl.l_len = 1; + /* Try a 1 byte lock in each possible position number */ for (fl.l_start = 0; fl.l_start < getpagesize()/sizeof(struct lguest_net); fl.l_start++) { + /* If we succeed, return the slot number. */ if (fcntl(netfd, F_SETLK, &fl) == 0) return fl.l_start; } errx(1, "No free slots in network file %s", filename); } +/* This function sets up the network file */ static void setup_net_file(const char *filename, struct device_list *devices) { int netfd; struct device *dev; + /* We don't use open_or_die() here: for friendliness we create the file + * if it doesn't already exist. */ netfd = open(filename, O_RDWR, 0); if (netfd < 0) { if (errno == ENOENT) { netfd = open(filename, O_RDWR|O_CREAT, 0600); if (netfd >= 0) { + /* If we succeeded, initialize the file with a + * blank page. */ char page[getpagesize()]; memset(page, 0, sizeof(page)); write(netfd, page, sizeof(page)); @@ -763,11 +1166,15 @@ static void setup_net_file(const char *filename, err(1, "cannot open net file '%s'", filename); } + /* We need 1 page, and the features indicate the slot to use and that + * no checksum is needed. We never touch this device again; it's + * between the Guests on the network, so we don't register input or + * output handlers. */ dev = new_device(devices, LGUEST_DEVICE_T_NET, 1, find_slot(netfd, filename)|LGUEST_NET_F_NOCSUM, -1, NULL, 0, NULL); - /* We overwrite the /dev/zero mapping with the actual file. */ + /* Map the shared file. */ if (mmap(dev->mem, getpagesize(), PROT_READ|PROT_WRITE, MAP_FIXED|MAP_SHARED, netfd, 0) != dev->mem) err(1, "could not mmap '%s'", filename); @@ -775,6 +1182,7 @@ static void setup_net_file(const char *filename, (void *)(dev->desc->pfn * getpagesize()), filename, dev->desc->features & ~LGUEST_NET_F_NOCSUM); } +/*:*/ static u32 str2ip(const char *ipaddr) { @@ -784,7 +1192,11 @@ static u32 str2ip(const char *ipaddr) return (byte[0] << 24) | (byte[1] << 16) | (byte[2] << 8) | byte[3]; } -/* adapted from libbridge */ +/* This code is "adapted" from libbridge: it attaches the Host end of the + * network device to the bridge device specified by the command line. + * + * This is yet another James Morris contribution (I'm an IP-level guy, so I + * dislike bridging), and I just try not to break it. */ static void add_to_bridge(int fd, const char *if_name, const char *br_name) { int ifidx; @@ -803,12 +1215,16 @@ static void add_to_bridge(int fd, const char *if_name, const char *br_name) err(1, "can't add %s to bridge %s", if_name, br_name); } +/* This sets up the Host end of the network device with an IP address, brings + * it up so packets will flow, the copies the MAC address into the hwaddr + * pointer (in practice, the Host's slot in the network device's memory). */ static void configure_device(int fd, const char *devname, u32 ipaddr, unsigned char hwaddr[6]) { struct ifreq ifr; struct sockaddr_in *sin = (struct sockaddr_in *)&ifr.ifr_addr; + /* Don't read these incantations. Just cut & paste them like I did! */ memset(&ifr, 0, sizeof(ifr)); strcpy(ifr.ifr_name, devname); sin->sin_family = AF_INET; @@ -819,12 +1235,19 @@ static void configure_device(int fd, const char *devname, u32 ipaddr, if (ioctl(fd, SIOCSIFFLAGS, &ifr) != 0) err(1, "Bringing interface %s up", devname); + /* SIOC stands for Socket I/O Control. G means Get (vs S for Set + * above). IF means Interface, and HWADDR is hardware address. + * Simple! */ if (ioctl(fd, SIOCGIFHWADDR, &ifr) != 0) err(1, "getting hw address for %s", devname); - memcpy(hwaddr, ifr.ifr_hwaddr.sa_data, 6); } +/*L:195 The other kind of network is a Host<->Guest network. This can either + * use briding or routing, but the principle is the same: it uses the "tun" + * device to inject packets into the Host as if they came in from a normal + * network card. We just shunt packets between the Guest and the tun + * device. */ static void setup_tun_net(const char *arg, struct device_list *devices) { struct device *dev; @@ -833,36 +1256,56 @@ static void setup_tun_net(const char *arg, struct device_list *devices) u32 ip; const char *br_name = NULL; + /* We open the /dev/net/tun device and tell it we want a tap device. A + * tap device is like a tun device, only somehow different. To tell + * the truth, I completely blundered my way through this code, but it + * works now! */ netfd = open_or_die("/dev/net/tun", O_RDWR); memset(&ifr, 0, sizeof(ifr)); ifr.ifr_flags = IFF_TAP | IFF_NO_PI; strcpy(ifr.ifr_name, "tap%d"); if (ioctl(netfd, TUNSETIFF, &ifr) != 0) err(1, "configuring /dev/net/tun"); + /* We don't need checksums calculated for packets coming in this + * device: trust us! */ ioctl(netfd, TUNSETNOCSUM, 1); - /* You will be peer 1: we should create enough jitter to randomize */ + /* We create the net device with 1 page, using the features field of + * the descriptor to tell the Guest it is in slot 1 (NET_PEERNUM), and + * that the device has fairly random timing. We do *not* specify + * LGUEST_NET_F_NOCSUM: these packets can reach the real world. + * + * We will put our MAC address is slot 0 for the Guest to see, so + * it will send packets to us using the key "peer_offset(0)": */ dev = new_device(devices, LGUEST_DEVICE_T_NET, 1, NET_PEERNUM|LGUEST_DEVICE_F_RANDOMNESS, netfd, handle_tun_input, peer_offset(0), handle_tun_output); + + /* We keep a flag which says whether we've seen packets come out from + * this network device. */ dev->priv = malloc(sizeof(bool)); *(bool *)dev->priv = false; + /* We need a socket to perform the magic network ioctls to bring up the + * tap interface, connect to the bridge etc. Any socket will do! */ ipfd = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP); if (ipfd < 0) err(1, "opening IP socket"); + /* If the command line was --tunnet=bridge: do bridging. */ if (!strncmp(BRIDGE_PFX, arg, strlen(BRIDGE_PFX))) { ip = INADDR_ANY; br_name = arg + strlen(BRIDGE_PFX); add_to_bridge(ipfd, ifr.ifr_name, br_name); - } else + } else /* It is an IP address to set up the device with */ ip = str2ip(arg); - /* We are peer 0, ie. first slot. */ + /* We are peer 0, ie. first slot, so we hand dev->mem to this routine + * to write the MAC address at the start of the device memory. */ configure_device(ipfd, ifr.ifr_name, ip, dev->mem); - /* Set "promisc" bit: we want every single packet. */ + /* Set "promisc" bit: we want every single packet if we're going to + * bridge to other machines (and otherwise it doesn't matter). */ *((u8 *)dev->mem) |= 0x1; close(ipfd); @@ -873,7 +1316,10 @@ static void setup_tun_net(const char *arg, struct device_list *devices) if (br_name) verbose("attached to bridge: %s\n", br_name); } +/* That's the end of device setup. */ +/*L:220 Finally we reach the core of the Launcher, which runs the Guest, serves + * its input and output, and finally, lays it to rest. */ static void __attribute__((noreturn)) run_guest(int lguest_fd, struct device_list *device_list) { @@ -885,20 +1331,37 @@ run_guest(int lguest_fd, struct device_list *device_list) /* We read from the /dev/lguest device to run the Guest. */ readval = read(lguest_fd, arr, sizeof(arr)); + /* The read can only really return sizeof(arr) (the Guest did a + * SEND_DMA to us), or an error. */ + + /* For a successful read, arr[0] is the address of the "struct + * lguest_dma", and arr[1] is the key the Guest sent to. */ if (readval == sizeof(arr)) { handle_output(lguest_fd, arr[0], arr[1], device_list); continue; + /* ENOENT means the Guest died. Reading tells us why. */ } else if (errno == ENOENT) { char reason[1024] = { 0 }; read(lguest_fd, reason, sizeof(reason)-1); errx(1, "%s", reason); + /* EAGAIN means the waker wanted us to look at some input. + * Anything else means a bug or incompatible change. */ } else if (errno != EAGAIN) err(1, "Running guest failed"); + + /* Service input, then unset the BREAK which releases + * the Waker. */ handle_input(lguest_fd, device_list); if (write(lguest_fd, args, sizeof(args)) < 0) err(1, "Resetting break"); } } +/* + * This is the end of the Launcher. + * + * But wait! We've seen I/O from the Launcher, and we've seen I/O from the + * Drivers. If we were to see the Host kernel I/O code, our understanding + * would be complete... :*/ static struct option opts[] = { { "verbose", 0, NULL, 'v' }, @@ -916,20 +1379,49 @@ static void usage(void) " vmlinux [args...]"); } +/*L:100 The Launcher code itself takes us out into userspace, that scary place + * where pointers run wild and free! Unfortunately, like most userspace + * programs, it's quite boring (which is why everyone like to hack on the + * kernel!). Perhaps if you make up an Lguest Drinking Game at this point, it + * will get you through this section. Or, maybe not. + * + * The Launcher binary sits up high, usually starting at address 0xB8000000. + * Everything below this is the "physical" memory for the Guest. For example, + * if the Guest were to write a "1" at physical address 0, we would see a "1" + * in the Launcher at "(int *)0". Guest physical == Launcher virtual. + * + * This can be tough to get your head around, but usually it just means that we + * don't need to do any conversion when the Guest gives us it's "physical" + * addresses. + */ int main(int argc, char *argv[]) { + /* Memory, top-level pagetable, code startpoint, PAGE_OFFSET and size + * of the (optional) initrd. */ unsigned long mem = 0, pgdir, start, page_offset, initrd_size = 0; + /* A temporary and the /dev/lguest file descriptor. */ int i, c, lguest_fd; + /* The list of Guest devices, based on command line arguments. */ struct device_list device_list; + /* The boot information for the Guest: at guest-physical address 0. */ void *boot = (void *)0; + /* If they specify an initrd file to load. */ const char *initrd_name = NULL; + /* First we initialize the device list. Since console and network + * device receive input from a file descriptor, we keep an fdset + * (infds) and the maximum fd number (max_infd) with the head of the + * list. We also keep a pointer to the last device, for easy appending + * to the list. */ device_list.max_infd = -1; device_list.dev = NULL; device_list.lastdev = &device_list.dev; FD_ZERO(&device_list.infds); - /* We need to know how much memory so we can allocate devices. */ + /* We need to know how much memory so we can set up the device + * descriptor and memory pages for the devices as we parse the command + * line. So we quickly look through the arguments to find the amount + * of memory now. */ for (i = 1; i < argc; i++) { if (argv[i][0] != '-') { mem = top = atoi(argv[i]) * 1024 * 1024; @@ -938,6 +1430,8 @@ int main(int argc, char *argv[]) break; } } + + /* The options are fairly straight-forward */ while ((c = getopt_long(argc, argv, "v", opts, NULL)) != EOF) { switch (c) { case 'v': @@ -960,42 +1454,59 @@ int main(int argc, char *argv[]) usage(); } } + /* After the other arguments we expect memory and kernel image name, + * followed by command line arguments for the kernel. */ if (optind + 2 > argc) usage(); - /* We need a console device */ + /* We always have a console device */ setup_console(&device_list); - /* First we map /dev/zero over all of guest-physical memory. */ + /* We start by mapping anonymous pages over all of guest-physical + * memory range. This fills it with 0, and ensures that the Guest + * won't be killed when it tries to access it. */ map_zeroed_pages(0, mem / getpagesize()); /* Now we load the kernel */ start = load_kernel(open_or_die(argv[optind+1], O_RDONLY), &page_offset); - /* Map the initrd image if requested */ + /* Map the initrd image if requested (at top of physical memory) */ if (initrd_name) { initrd_size = load_initrd(initrd_name, mem); + /* These are the location in the Linux boot header where the + * start and size of the initrd are expected to be found. */ *(unsigned long *)(boot+0x218) = mem - initrd_size; *(unsigned long *)(boot+0x21c) = initrd_size; + /* The bootloader type 0xFF means "unknown"; that's OK. */ *(unsigned char *)(boot+0x210) = 0xFF; } - /* Set up the initial linar pagetables. */ + /* Set up the initial linear pagetables, starting below the initrd. */ pgdir = setup_pagetables(mem, initrd_size, page_offset); - /* E820 memory map: ours is a simple, single region. */ + /* The Linux boot header contains an "E820" memory map: ours is a + * simple, single region. */ *(char*)(boot+E820NR) = 1; *((struct e820entry *)(boot+E820MAP)) = ((struct e820entry) { 0, mem, E820_RAM }); - /* Command line pointer and command line (at 4096) */ + /* The boot header contains a command line pointer: we put the command + * line after the boot header (at address 4096) */ *(void **)(boot + 0x228) = boot + 4096; concat(boot + 4096, argv+optind+2); - /* Paravirt type: 1 == lguest */ + + /* The guest type value of "1" tells the Guest it's under lguest. */ *(int *)(boot + 0x23c) = 1; + /* We tell the kernel to initialize the Guest: this returns the open + * /dev/lguest file descriptor. */ lguest_fd = tell_kernel(pgdir, start, page_offset); + + /* We fork off a child process, which wakes the Launcher whenever one + * of the input file descriptors needs attention. Otherwise we would + * run the Guest until it tries to output something. */ waker_fd = setup_waker(lguest_fd, &device_list); + /* Finally, run the Guest. This doesn't return. */ run_guest(lguest_fd, &device_list); } diff --git a/drivers/lguest/core.c b/drivers/lguest/core.c index 2cea0c80c992..1eb05f9a56b6 100644 --- a/drivers/lguest/core.c +++ b/drivers/lguest/core.c @@ -208,24 +208,39 @@ static int emulate_insn(struct lguest *lg) return 1; } +/*L:305 + * Dealing With Guest Memory. + * + * When the Guest gives us (what it thinks is) a physical address, we can use + * the normal copy_from_user() & copy_to_user() on that address: remember, + * Guest physical == Launcher virtual. + * + * But we can't trust the Guest: it might be trying to access the Launcher + * code. We have to check that the range is below the pfn_limit the Launcher + * gave us. We have to make sure that addr + len doesn't give us a false + * positive by overflowing, too. */ int lguest_address_ok(const struct lguest *lg, unsigned long addr, unsigned long len) { return (addr+len) / PAGE_SIZE < lg->pfn_limit && (addr+len >= addr); } -/* Just like get_user, but don't let guest access lguest binary. */ +/* This is a convenient routine to get a 32-bit value from the Guest (a very + * common operation). Here we can see how useful the kill_lguest() routine we + * met in the Launcher can be: we return a random value (0) instead of needing + * to return an error. */ u32 lgread_u32(struct lguest *lg, unsigned long addr) { u32 val = 0; - /* Don't let them access lguest binary */ + /* Don't let them access lguest binary. */ if (!lguest_address_ok(lg, addr, sizeof(val)) || get_user(val, (u32 __user *)addr) != 0) kill_guest(lg, "bad read address %#lx", addr); return val; } +/* Same thing for writing a value. */ void lgwrite_u32(struct lguest *lg, unsigned long addr, u32 val) { if (!lguest_address_ok(lg, addr, sizeof(val)) @@ -233,6 +248,9 @@ void lgwrite_u32(struct lguest *lg, unsigned long addr, u32 val) kill_guest(lg, "bad write address %#lx", addr); } +/* This routine is more generic, and copies a range of Guest bytes into a + * buffer. If the copy_from_user() fails, we fill the buffer with zeroes, so + * the caller doesn't end up using uninitialized kernel memory. */ void lgread(struct lguest *lg, void *b, unsigned long addr, unsigned bytes) { if (!lguest_address_ok(lg, addr, bytes) @@ -243,6 +261,7 @@ void lgread(struct lguest *lg, void *b, unsigned long addr, unsigned bytes) } } +/* Similarly, our generic routine to copy into a range of Guest bytes. */ void lgwrite(struct lguest *lg, unsigned long addr, const void *b, unsigned bytes) { @@ -250,6 +269,7 @@ void lgwrite(struct lguest *lg, unsigned long addr, const void *b, || copy_to_user((void __user *)addr, b, bytes) != 0) kill_guest(lg, "bad write address %#lx len %u", addr, bytes); } +/* (end of memory access helper routines) :*/ static void set_ts(void) { diff --git a/drivers/lguest/io.c b/drivers/lguest/io.c index d2f02f0653ca..da288128e44f 100644 --- a/drivers/lguest/io.c +++ b/drivers/lguest/io.c @@ -27,8 +27,36 @@ #include #include "lg.h" +/*L:300 + * I/O + * + * Getting data in and out of the Guest is quite an art. There are numerous + * ways to do it, and they all suck differently. We try to keep things fairly + * close to "real" hardware so our Guest's drivers don't look like an alien + * visitation in the middle of the Linux code, and yet make sure that Guests + * can talk directly to other Guests, not just the Launcher. + * + * To do this, the Guest gives us a key when it binds or sends DMA buffers. + * The key corresponds to a "physical" address inside the Guest (ie. a virtual + * address inside the Launcher process). We don't, however, use this key + * directly. + * + * We want Guests which share memory to be able to DMA to each other: two + * Launchers can mmap memory the same file, then the Guests can communicate. + * Fortunately, the futex code provides us with a way to get a "union + * futex_key" corresponding to the memory lying at a virtual address: if the + * two processes share memory, the "union futex_key" for that memory will match + * even if the memory is mapped at different addresses in each. So we always + * convert the keys to "union futex_key"s to compare them. + * + * Before we dive into this though, we need to look at another set of helper + * routines used throughout the Host kernel code to access Guest memory. + :*/ static struct list_head dma_hash[61]; +/* An unfortunate side effect of the Linux double-linked list implementation is + * that there's no good way to statically initialize an array of linked + * lists. */ void lguest_io_init(void) { unsigned int i; @@ -60,6 +88,19 @@ kill: return 0; } +/*L:330 This is our hash function, using the wonderful Jenkins hash. + * + * The futex key is a union with three parts: an unsigned long word, a pointer, + * and an int "offset". We could use jhash_2words() which takes three u32s. + * (Ok, the hash functions are great: the naming sucks though). + * + * It's nice to be portable to 64-bit platforms, so we use the more generic + * jhash2(), which takes an array of u32, the number of u32s, and an initial + * u32 to roll in. This is uglier, but breaks down to almost the same code on + * 32-bit platforms like this one. + * + * We want a position in the array, so we modulo ARRAY_SIZE(dma_hash) (ie. 61). + */ static unsigned int hash(const union futex_key *key) { return jhash2((u32*)&key->both.word, @@ -68,6 +109,9 @@ static unsigned int hash(const union futex_key *key) % ARRAY_SIZE(dma_hash); } +/* This is a convenience routine to compare two keys. It's a much bemoaned C + * weakness that it doesn't allow '==' on structures or unions, so we have to + * open-code it like this. */ static inline int key_eq(const union futex_key *a, const union futex_key *b) { return (a->both.word == b->both.word @@ -75,22 +119,36 @@ static inline int key_eq(const union futex_key *a, const union futex_key *b) && a->both.offset == b->both.offset); } -/* Must hold read lock on dmainfo owner's current->mm->mmap_sem */ +/*L:360 OK, when we need to actually free up a Guest's DMA array we do several + * things, so we have a convenient function to do it. + * + * The caller must hold a read lock on dmainfo owner's current->mm->mmap_sem + * for the drop_futex_key_refs(). */ static void unlink_dma(struct lguest_dma_info *dmainfo) { + /* You locked this too, right? */ BUG_ON(!mutex_is_locked(&lguest_lock)); + /* This is how we know that the entry is free. */ dmainfo->interrupt = 0; + /* Remove it from the hash table. */ list_del(&dmainfo->list); + /* Drop the references we were holding (to the inode or mm). */ drop_futex_key_refs(&dmainfo->key); } +/*L:350 This is the routine which we call when the Guest asks to unregister a + * DMA array attached to a given key. Returns true if the array was found. */ static int unbind_dma(struct lguest *lg, const union futex_key *key, unsigned long dmas) { int i, ret = 0; + /* We don't bother with the hash table, just look through all this + * Guest's DMA arrays. */ for (i = 0; i < LGUEST_MAX_DMA; i++) { + /* In theory it could have more than one array on the same key, + * or one array on multiple keys, so we check both */ if (key_eq(key, &lg->dma[i].key) && dmas == lg->dma[i].dmas) { unlink_dma(&lg->dma[i]); ret = 1; @@ -100,51 +158,91 @@ static int unbind_dma(struct lguest *lg, return ret; } +/*L:340 BIND_DMA: this is the hypercall which sets up an array of "struct + * lguest_dma" for receiving I/O. + * + * The Guest wants to bind an array of "struct lguest_dma"s to a particular key + * to receive input. This only happens when the Guest is setting up a new + * device, so it doesn't have to be very fast. + * + * It returns 1 on a successful registration (it can fail if we hit the limit + * of registrations for this Guest). + */ int bind_dma(struct lguest *lg, unsigned long ukey, unsigned long dmas, u16 numdmas, u8 interrupt) { unsigned int i; int ret = 0; union futex_key key; + /* Futex code needs the mmap_sem. */ struct rw_semaphore *fshared = ¤t->mm->mmap_sem; + /* Invalid interrupt? (We could kill the guest here). */ if (interrupt >= LGUEST_IRQS) return 0; + /* We need to grab the Big Lguest Lock, because other Guests may be + * trying to look through this Guest's DMAs to send something while + * we're doing this. */ mutex_lock(&lguest_lock); down_read(fshared); if (get_futex_key((u32 __user *)ukey, fshared, &key) != 0) { kill_guest(lg, "bad dma key %#lx", ukey); goto unlock; } + + /* We want to keep this key valid once we drop mmap_sem, so we have to + * hold a reference. */ get_futex_key_refs(&key); + /* If the Guest specified an interrupt of 0, that means they want to + * unregister this array of "struct lguest_dma"s. */ if (interrupt == 0) ret = unbind_dma(lg, &key, dmas); else { + /* Look through this Guest's dma array for an unused entry. */ for (i = 0; i < LGUEST_MAX_DMA; i++) { + /* If the interrupt is non-zero, the entry is already + * used. */ if (lg->dma[i].interrupt) continue; + /* OK, a free one! Fill on our details. */ lg->dma[i].dmas = dmas; lg->dma[i].num_dmas = numdmas; lg->dma[i].next_dma = 0; lg->dma[i].key = key; lg->dma[i].guestid = lg->guestid; lg->dma[i].interrupt = interrupt; + + /* Now we add it to the hash table: the position + * depends on the futex key that we got. */ list_add(&lg->dma[i].list, &dma_hash[hash(&key)]); + /* Success! */ ret = 1; goto unlock; } } + /* If we didn't find a slot to put the key in, drop the reference + * again. */ drop_futex_key_refs(&key); unlock: + /* Unlock and out. */ up_read(fshared); mutex_unlock(&lguest_lock); return ret; } -/* lgread from another guest */ +/*L:385 Note that our routines to access a different Guest's memory are called + * lgread_other() and lgwrite_other(): these names emphasize that they are only + * used when the Guest is *not* the current Guest. + * + * The interface for copying from another process's memory is called + * access_process_vm(), with a final argument of 0 for a read, and 1 for a + * write. + * + * We need lgread_other() to read the destination Guest's "struct lguest_dma" + * array. */ static int lgread_other(struct lguest *lg, void *buf, u32 addr, unsigned bytes) { @@ -157,7 +255,8 @@ static int lgread_other(struct lguest *lg, return 1; } -/* lgwrite to another guest */ +/* "lgwrite()" to another Guest: used to update the destination "used_len" once + * we've transferred data into the buffer. */ static int lgwrite_other(struct lguest *lg, u32 addr, const void *buf, unsigned bytes) { @@ -170,6 +269,15 @@ static int lgwrite_other(struct lguest *lg, u32 addr, return 1; } +/*L:400 This is the generic engine which copies from a source "struct + * lguest_dma" from this Guest into another Guest's "struct lguest_dma". The + * destination Guest's pages have already been mapped, as contained in the + * pages array. + * + * If you're wondering if there's a nice "copy from one process to another" + * routine, so was I. But Linux isn't really set up to copy between two + * unrelated processes, so we have to write it ourselves. + */ static u32 copy_data(struct lguest *srclg, const struct lguest_dma *src, const struct lguest_dma *dst, @@ -178,33 +286,59 @@ static u32 copy_data(struct lguest *srclg, unsigned int totlen, si, di, srcoff, dstoff; void *maddr = NULL; + /* We return the total length transferred. */ totlen = 0; + + /* We keep indexes into the source and destination "struct lguest_dma", + * and an offset within each region. */ si = di = 0; srcoff = dstoff = 0; + + /* We loop until the source or destination is exhausted. */ while (si < LGUEST_MAX_DMA_SECTIONS && src->len[si] && di < LGUEST_MAX_DMA_SECTIONS && dst->len[di]) { + /* We can only transfer the rest of the src buffer, or as much + * as will fit into the destination buffer. */ u32 len = min(src->len[si] - srcoff, dst->len[di] - dstoff); + /* For systems using "highmem" we need to use kmap() to access + * the page we want. We often use the same page over and over, + * so rather than kmap() it on every loop, we set the maddr + * pointer to NULL when we need to move to the next + * destination page. */ if (!maddr) maddr = kmap(pages[di]); - /* FIXME: This is not completely portable, since - archs do different things for copy_to_user_page. */ + /* Copy directly from (this Guest's) source address to the + * destination Guest's kmap()ed buffer. Note that maddr points + * to the start of the page: we need to add the offset of the + * destination address and offset within the buffer. */ + + /* FIXME: This is not completely portable. I looked at + * copy_to_user_page(), and some arch's seem to need special + * flushes. x86 is fine. */ if (copy_from_user(maddr + (dst->addr[di] + dstoff)%PAGE_SIZE, (void __user *)src->addr[si], len) != 0) { + /* If a copy failed, it's the source's fault. */ kill_guest(srclg, "bad address in sending DMA"); totlen = 0; break; } + /* Increment the total and src & dst offsets */ totlen += len; srcoff += len; dstoff += len; + + /* Presumably we reached the end of the src or dest buffers: */ if (srcoff == src->len[si]) { + /* Move to the next buffer at offset 0 */ si++; srcoff = 0; } if (dstoff == dst->len[di]) { + /* We need to unmap that destination page and reset + * maddr ready for the next one. */ kunmap(pages[di]); maddr = NULL; di++; @@ -212,13 +346,15 @@ static u32 copy_data(struct lguest *srclg, } } + /* If we still had a page mapped at the end, unmap now. */ if (maddr) kunmap(pages[di]); return totlen; } -/* Src is us, ie. current. */ +/*L:390 This is how we transfer a "struct lguest_dma" from the source Guest + * (the current Guest which called SEND_DMA) to another Guest. */ static u32 do_dma(struct lguest *srclg, const struct lguest_dma *src, struct lguest *dstlg, const struct lguest_dma *dst) { @@ -226,23 +362,31 @@ static u32 do_dma(struct lguest *srclg, const struct lguest_dma *src, u32 ret; struct page *pages[LGUEST_MAX_DMA_SECTIONS]; + /* We check that both source and destination "struct lguest_dma"s are + * within the bounds of the source and destination Guests */ if (!check_dma_list(dstlg, dst) || !check_dma_list(srclg, src)) return 0; - /* First get the destination pages */ + /* We need to map the pages which correspond to each parts of + * destination buffer. */ for (i = 0; i < LGUEST_MAX_DMA_SECTIONS; i++) { if (dst->len[i] == 0) break; + /* get_user_pages() is a complicated function, especially since + * we only want a single page. But it works, and returns the + * number of pages. Note that we're holding the destination's + * mmap_sem, as get_user_pages() requires. */ if (get_user_pages(dstlg->tsk, dstlg->mm, dst->addr[i], 1, 1, 1, pages+i, NULL) != 1) { + /* This means the destination gave us a bogus buffer */ kill_guest(dstlg, "Error mapping DMA pages"); ret = 0; goto drop_pages; } } - /* Now copy until we run out of src or dst. */ + /* Now copy the data until we run out of src or dst. */ ret = copy_data(srclg, src, dst, pages); drop_pages: @@ -251,6 +395,11 @@ drop_pages: return ret; } +/*L:380 Transferring data from one Guest to another is not as simple as I'd + * like. We've found the "struct lguest_dma_info" bound to the same address as + * the send, we need to copy into it. + * + * This function returns true if the destination array was empty. */ static int dma_transfer(struct lguest *srclg, unsigned long udma, struct lguest_dma_info *dst) @@ -259,15 +408,23 @@ static int dma_transfer(struct lguest *srclg, struct lguest *dstlg; u32 i, dma = 0; + /* From the "struct lguest_dma_info" we found in the hash, grab the + * Guest. */ dstlg = &lguests[dst->guestid]; - /* Get our dma list. */ + /* Read in the source "struct lguest_dma" handed to SEND_DMA. */ lgread(srclg, &src_dma, udma, sizeof(src_dma)); - /* We can't deadlock against them dmaing to us, because this - * is all under the lguest_lock. */ + /* We need the destination's mmap_sem, and we already hold the source's + * mmap_sem for the futex key lookup. Normally this would suggest that + * we could deadlock if the destination Guest was trying to send to + * this source Guest at the same time, which is another reason that all + * I/O is done under the big lguest_lock. */ down_read(&dstlg->mm->mmap_sem); + /* Look through the destination DMA array for an available buffer. */ for (i = 0; i < dst->num_dmas; i++) { + /* We keep a "next_dma" pointer which often helps us avoid + * looking at lots of previously-filled entries. */ dma = (dst->next_dma + i) % dst->num_dmas; if (!lgread_other(dstlg, &dst_dma, dst->dmas + dma * sizeof(struct lguest_dma), @@ -277,30 +434,46 @@ static int dma_transfer(struct lguest *srclg, if (!dst_dma.used_len) break; } + + /* If we found a buffer, we do the actual data copy. */ if (i != dst->num_dmas) { unsigned long used_lenp; unsigned int ret; ret = do_dma(srclg, &src_dma, dstlg, &dst_dma); - /* Put used length in src. */ + /* Put used length in the source "struct lguest_dma"'s used_len + * field. It's a little tricky to figure out where that is, + * though. */ lgwrite_u32(srclg, udma+offsetof(struct lguest_dma, used_len), ret); + /* Tranferring 0 bytes is OK if the source buffer was empty. */ if (ret == 0 && src_dma.len[0] != 0) goto fail; - /* Make sure destination sees contents before length. */ + /* The destination Guest might be running on a different CPU: + * we have to make sure that it will see the "used_len" field + * change to non-zero *after* it sees the data we copied into + * the buffer. Hence a write memory barrier. */ wmb(); + /* Figuring out where the destination's used_len field for this + * "struct lguest_dma" in the array is also a little ugly. */ used_lenp = dst->dmas + dma * sizeof(struct lguest_dma) + offsetof(struct lguest_dma, used_len); lgwrite_other(dstlg, used_lenp, &ret, sizeof(ret)); + /* Move the cursor for next time. */ dst->next_dma++; } up_read(&dstlg->mm->mmap_sem); - /* Do this last so dst doesn't simply sleep on lock. */ + /* We trigger the destination interrupt, even if the destination was + * empty and we didn't transfer anything: this gives them a chance to + * wake up and refill. */ set_bit(dst->interrupt, dstlg->irqs_pending); + /* Wake up the destination process. */ wake_up_process(dstlg->tsk); + /* If we passed the last "struct lguest_dma", the receive had no + * buffers left. */ return i == dst->num_dmas; fail: @@ -308,6 +481,8 @@ fail: return 0; } +/*L:370 This is the counter-side to the BIND_DMA hypercall; the SEND_DMA + * hypercall. We find out who's listening, and send to them. */ void send_dma(struct lguest *lg, unsigned long ukey, unsigned long udma) { union futex_key key; @@ -317,31 +492,43 @@ void send_dma(struct lguest *lg, unsigned long ukey, unsigned long udma) again: mutex_lock(&lguest_lock); down_read(fshared); + /* Get the futex key for the key the Guest gave us */ if (get_futex_key((u32 __user *)ukey, fshared, &key) != 0) { kill_guest(lg, "bad sending DMA key"); goto unlock; } - /* Shared mapping? Look for other guests... */ + /* Since the key must be a multiple of 4, the futex key uses the lower + * bit of the "offset" field (which would always be 0) to indicate a + * mapping which is shared with other processes (ie. Guests). */ if (key.shared.offset & 1) { struct lguest_dma_info *i; + /* Look through the hash for other Guests. */ list_for_each_entry(i, &dma_hash[hash(&key)], list) { + /* Don't send to ourselves. */ if (i->guestid == lg->guestid) continue; if (!key_eq(&key, &i->key)) continue; + /* If dma_transfer() tells us the destination has no + * available buffers, we increment "empty". */ empty += dma_transfer(lg, udma, i); break; } + /* If the destination is empty, we release our locks and + * give the destination Guest a brief chance to restock. */ if (empty == 1) { /* Give any recipients one chance to restock. */ up_read(¤t->mm->mmap_sem); mutex_unlock(&lguest_lock); + /* Next time, we won't try again. */ empty++; goto again; } } else { - /* Private mapping: tell our userspace. */ + /* Private mapping: Guest is sending to its Launcher. We set + * the "dma_is_pending" flag so that the main loop will exit + * and the Launcher's read() from /dev/lguest will return. */ lg->dma_is_pending = 1; lg->pending_dma = udma; lg->pending_key = ukey; @@ -350,6 +537,7 @@ unlock: up_read(fshared); mutex_unlock(&lguest_lock); } +/*:*/ void release_all_dma(struct lguest *lg) { @@ -365,7 +553,8 @@ void release_all_dma(struct lguest *lg) up_read(&lg->mm->mmap_sem); } -/* Userspace wants a dma buffer from this guest. */ +/*L:320 This routine looks for a DMA buffer registered by the Guest on the + * given key (using the BIND_DMA hypercall). */ unsigned long get_dma_buffer(struct lguest *lg, unsigned long ukey, unsigned long *interrupt) { @@ -374,15 +563,29 @@ unsigned long get_dma_buffer(struct lguest *lg, struct lguest_dma_info *i; struct rw_semaphore *fshared = ¤t->mm->mmap_sem; + /* Take the Big Lguest Lock to stop other Guests sending this Guest DMA + * at the same time. */ mutex_lock(&lguest_lock); + /* To match between Guests sharing the same underlying memory we steal + * code from the futex infrastructure. This requires that we hold the + * "mmap_sem" for our process (the Launcher), and pass it to the futex + * code. */ down_read(fshared); + + /* This can fail if it's not a valid address, or if the address is not + * divisible by 4 (the futex code needs that, we don't really). */ if (get_futex_key((u32 __user *)ukey, fshared, &key) != 0) { kill_guest(lg, "bad registered DMA buffer"); goto unlock; } + /* Search the hash table for matching entries (the Launcher can only + * send to its own Guest for the moment, so the entry must be for this + * Guest) */ list_for_each_entry(i, &dma_hash[hash(&key)], list) { if (key_eq(&key, &i->key) && i->guestid == lg->guestid) { unsigned int j; + /* Look through the registered DMA array for an + * available buffer. */ for (j = 0; j < i->num_dmas; j++) { struct lguest_dma dma; @@ -391,6 +594,8 @@ unsigned long get_dma_buffer(struct lguest *lg, if (dma.used_len == 0) break; } + /* Store the interrupt the Guest wants when the buffer + * is used. */ *interrupt = i->interrupt; break; } @@ -400,4 +605,12 @@ unlock: mutex_unlock(&lguest_lock); return ret; } +/*:*/ +/*L:410 This really has completed the Launcher. Not only have we now finished + * the longest chapter in our journey, but this also means we are over halfway + * through! + * + * Enough prevaricating around the bush: it is time for us to dive into the + * core of the Host, in "make Host". + */ diff --git a/drivers/lguest/lg.h b/drivers/lguest/lg.h index 3e2ddfbc816e..3b9dc123a7df 100644 --- a/drivers/lguest/lg.h +++ b/drivers/lguest/lg.h @@ -244,6 +244,30 @@ unsigned long get_dma_buffer(struct lguest *lg, unsigned long key, /* hypercalls.c: */ void do_hypercalls(struct lguest *lg); +/*L:035 + * Let's step aside for the moment, to study one important routine that's used + * widely in the Host code. + * + * There are many cases where the Guest does something invalid, like pass crap + * to a hypercall. Since only the Guest kernel can make hypercalls, it's quite + * acceptable to simply terminate the Guest and give the Launcher a nicely + * formatted reason. It's also simpler for the Guest itself, which doesn't + * need to check most hypercalls for "success"; if you're still running, it + * succeeded. + * + * Once this is called, the Guest will never run again, so most Host code can + * call this then continue as if nothing had happened. This means many + * functions don't have to explicitly return an error code, which keeps the + * code simple. + * + * It also means that this can be called more than once: only the first one is + * remembered. The only trick is that we still need to kill the Guest even if + * we can't allocate memory to store the reason. Linux has a neat way of + * packing error codes into invalid pointers, so we use that here. + * + * Like any macro which uses an "if", it is safely wrapped in a run-once "do { + * } while(0)". + */ #define kill_guest(lg, fmt...) \ do { \ if (!(lg)->dead) { \ @@ -252,6 +276,7 @@ do { \ (lg)->dead = ERR_PTR(-ENOMEM); \ } \ } while(0) +/* (End of aside) :*/ static inline unsigned long guest_pa(struct lguest *lg, unsigned long vaddr) { diff --git a/drivers/lguest/lguest_user.c b/drivers/lguest/lguest_user.c index 6ae86f20ce3d..80d1b58c7698 100644 --- a/drivers/lguest/lguest_user.c +++ b/drivers/lguest/lguest_user.c @@ -9,33 +9,62 @@ #include #include "lg.h" +/*L:030 setup_regs() doesn't really belong in this file, but it gives us an + * early glimpse deeper into the Host so it's worth having here. + * + * Most of the Guest's registers are left alone: we used get_zeroed_page() to + * allocate the structure, so they will be 0. */ static void setup_regs(struct lguest_regs *regs, unsigned long start) { - /* Write out stack in format lguest expects, so we can switch to it. */ + /* There are four "segment" registers which the Guest needs to boot: + * The "code segment" register (cs) refers to the kernel code segment + * __KERNEL_CS, and the "data", "extra" and "stack" segment registers + * refer to the kernel data segment __KERNEL_DS. + * + * The privilege level is packed into the lower bits. The Guest runs + * at privilege level 1 (GUEST_PL).*/ regs->ds = regs->es = regs->ss = __KERNEL_DS|GUEST_PL; regs->cs = __KERNEL_CS|GUEST_PL; - regs->eflags = 0x202; /* Interrupts enabled. */ + + /* The "eflags" register contains miscellaneous flags. Bit 1 (0x002) + * is supposed to always be "1". Bit 9 (0x200) controls whether + * interrupts are enabled. We always leave interrupts enabled while + * running the Guest. */ + regs->eflags = 0x202; + + /* The "Extended Instruction Pointer" register says where the Guest is + * running. */ regs->eip = start; - /* esi points to our boot information (physical address 0) */ + + /* %esi points to our boot information, at physical address 0, so don't + * touch it. */ } -/* + addr */ +/*L:310 To send DMA into the Guest, the Launcher needs to be able to ask for a + * DMA buffer. This is done by writing LHREQ_GETDMA and the key to + * /dev/lguest. */ static long user_get_dma(struct lguest *lg, const u32 __user *input) { unsigned long key, udma, irq; + /* Fetch the key they wrote to us. */ if (get_user(key, input) != 0) return -EFAULT; + /* Look for a free Guest DMA buffer bound to that key. */ udma = get_dma_buffer(lg, key, &irq); if (!udma) return -ENOENT; - /* We put irq number in udma->used_len. */ + /* We need to tell the Launcher what interrupt the Guest expects after + * the buffer is filled. We stash it in udma->used_len. */ lgwrite_u32(lg, udma + offsetof(struct lguest_dma, used_len), irq); + + /* The (guest-physical) address of the DMA buffer is returned from + * the write(). */ return udma; } -/* To force the Guest to stop running and return to the Launcher, the +/*L:315 To force the Guest to stop running and return to the Launcher, the * Waker sets writes LHREQ_BREAK and the value "1" to /dev/lguest. The * Launcher then writes LHREQ_BREAK and "0" to release the Waker. */ static int break_guest_out(struct lguest *lg, const u32 __user *input) @@ -59,7 +88,8 @@ static int break_guest_out(struct lguest *lg, const u32 __user *input) } } -/* + irq */ +/*L:050 Sending an interrupt is done by writing LHREQ_IRQ and an interrupt + * number to /dev/lguest. */ static int user_send_irq(struct lguest *lg, const u32 __user *input) { u32 irq; @@ -68,14 +98,19 @@ static int user_send_irq(struct lguest *lg, const u32 __user *input) return -EFAULT; if (irq >= LGUEST_IRQS) return -EINVAL; + /* Next time the Guest runs, the core code will see if it can deliver + * this interrupt. */ set_bit(irq, lg->irqs_pending); return 0; } +/*L:040 Once our Guest is initialized, the Launcher makes it run by reading + * from /dev/lguest. */ static ssize_t read(struct file *file, char __user *user, size_t size,loff_t*o) { struct lguest *lg = file->private_data; + /* You must write LHREQ_INITIALIZE first! */ if (!lg) return -EINVAL; @@ -83,27 +118,52 @@ static ssize_t read(struct file *file, char __user *user, size_t size,loff_t*o) if (current != lg->tsk) return -EPERM; + /* If the guest is already dead, we indicate why */ if (lg->dead) { size_t len; + /* lg->dead either contains an error code, or a string. */ if (IS_ERR(lg->dead)) return PTR_ERR(lg->dead); + /* We can only return as much as the buffer they read with. */ len = min(size, strlen(lg->dead)+1); if (copy_to_user(user, lg->dead, len) != 0) return -EFAULT; return len; } + /* If we returned from read() last time because the Guest sent DMA, + * clear the flag. */ if (lg->dma_is_pending) lg->dma_is_pending = 0; + /* Run the Guest until something interesting happens. */ return run_guest(lg, (unsigned long __user *)user); } -/* Take: pfnlimit, pgdir, start, pageoffset. */ +/*L:020 The initialization write supplies 4 32-bit values (in addition to the + * 32-bit LHREQ_INITIALIZE value). These are: + * + * pfnlimit: The highest (Guest-physical) page number the Guest should be + * allowed to access. The Launcher has to live in Guest memory, so it sets + * this to ensure the Guest can't reach it. + * + * pgdir: The (Guest-physical) address of the top of the initial Guest + * pagetables (which are set up by the Launcher). + * + * start: The first instruction to execute ("eip" in x86-speak). + * + * page_offset: The PAGE_OFFSET constant in the Guest kernel. We should + * probably wean the code off this, but it's a very useful constant! Any + * address above this is within the Guest kernel, and any kernel address can + * quickly converted from physical to virtual by adding PAGE_OFFSET. It's + * 0xC0000000 (3G) by default, but it's configurable at kernel build time. + */ static int initialize(struct file *file, const u32 __user *input) { + /* "struct lguest" contains everything we (the Host) know about a + * Guest. */ struct lguest *lg; int err, i; u32 args[4]; @@ -111,7 +171,7 @@ static int initialize(struct file *file, const u32 __user *input) /* We grab the Big Lguest lock, which protects the global array * "lguests" and multiple simultaneous initializations. */ mutex_lock(&lguest_lock); - + /* You can't initialize twice! Close the device and start again... */ if (file->private_data) { err = -EBUSY; goto unlock; @@ -122,37 +182,70 @@ static int initialize(struct file *file, const u32 __user *input) goto unlock; } + /* Find an unused guest. */ i = find_free_guest(); if (i < 0) { err = -ENOSPC; goto unlock; } + /* OK, we have an index into the "lguest" array: "lg" is a convenient + * pointer. */ lg = &lguests[i]; + + /* Populate the easy fields of our "struct lguest" */ lg->guestid = i; lg->pfn_limit = args[0]; lg->page_offset = args[3]; + + /* We need a complete page for the Guest registers: they are accessible + * to the Guest and we can only grant it access to whole pages. */ lg->regs_page = get_zeroed_page(GFP_KERNEL); if (!lg->regs_page) { err = -ENOMEM; goto release_guest; } + /* We actually put the registers at the bottom of the page. */ lg->regs = (void *)lg->regs_page + PAGE_SIZE - sizeof(*lg->regs); + /* Initialize the Guest's shadow page tables, using the toplevel + * address the Launcher gave us. This allocates memory, so can + * fail. */ err = init_guest_pagetable(lg, args[1]); if (err) goto free_regs; + /* Now we initialize the Guest's registers, handing it the start + * address. */ setup_regs(lg->regs, args[2]); + + /* There are a couple of GDT entries the Guest expects when first + * booting. */ setup_guest_gdt(lg); + + /* The timer for lguest's clock needs initialization. */ init_clockdev(lg); + + /* We keep a pointer to the Launcher task (ie. current task) for when + * other Guests want to wake this one (inter-Guest I/O). */ lg->tsk = current; + /* We need to keep a pointer to the Launcher's memory map, because if + * the Launcher dies we need to clean it up. If we don't keep a + * reference, it is destroyed before close() is called. */ lg->mm = get_task_mm(lg->tsk); + + /* Initialize the queue for the waker to wait on */ init_waitqueue_head(&lg->break_wq); + + /* We remember which CPU's pages this Guest used last, for optimization + * when the same Guest runs on the same CPU twice. */ lg->last_pages = NULL; + + /* We keep our "struct lguest" in the file's private_data. */ file->private_data = lg; mutex_unlock(&lguest_lock); + /* And because this is a write() call, we return the length used. */ return sizeof(args); free_regs: @@ -164,9 +257,15 @@ unlock: return err; } +/*L:010 The first operation the Launcher does must be a write. All writes + * start with a 32 bit number: for the first write this must be + * LHREQ_INITIALIZE to set up the Guest. After that the Launcher can use + * writes of other values to get DMA buffers and send interrupts. */ static ssize_t write(struct file *file, const char __user *input, size_t size, loff_t *off) { + /* Once the guest is initialized, we hold the "struct lguest" in the + * file private data. */ struct lguest *lg = file->private_data; u32 req; @@ -174,8 +273,11 @@ static ssize_t write(struct file *file, const char __user *input, return -EFAULT; input += sizeof(req); + /* If you haven't initialized, you must do that first. */ if (req != LHREQ_INITIALIZE && !lg) return -EINVAL; + + /* Once the Guest is dead, all you can do is read() why it died. */ if (lg && lg->dead) return -ENOENT; @@ -197,33 +299,72 @@ static ssize_t write(struct file *file, const char __user *input, } } +/*L:060 The final piece of interface code is the close() routine. It reverses + * everything done in initialize(). This is usually called because the + * Launcher exited. + * + * Note that the close routine returns 0 or a negative error number: it can't + * really fail, but it can whine. I blame Sun for this wart, and K&R C for + * letting them do it. :*/ static int close(struct inode *inode, struct file *file) { struct lguest *lg = file->private_data; + /* If we never successfully initialized, there's nothing to clean up */ if (!lg) return 0; + /* We need the big lock, to protect from inter-guest I/O and other + * Launchers initializing guests. */ mutex_lock(&lguest_lock); /* Cancels the hrtimer set via LHCALL_SET_CLOCKEVENT. */ hrtimer_cancel(&lg->hrt); + /* Free any DMA buffers the Guest had bound. */ release_all_dma(lg); + /* Free up the shadow page tables for the Guest. */ free_guest_pagetable(lg); + /* Now all the memory cleanups are done, it's safe to release the + * Launcher's memory management structure. */ mmput(lg->mm); + /* If lg->dead doesn't contain an error code it will be NULL or a + * kmalloc()ed string, either of which is ok to hand to kfree(). */ if (!IS_ERR(lg->dead)) kfree(lg->dead); + /* We can free up the register page we allocated. */ free_page(lg->regs_page); + /* We clear the entire structure, which also marks it as free for the + * next user. */ memset(lg, 0, sizeof(*lg)); + /* Release lock and exit. */ mutex_unlock(&lguest_lock); + return 0; } +/*L:000 + * Welcome to our journey through the Launcher! + * + * The Launcher is the Host userspace program which sets up, runs and services + * the Guest. In fact, many comments in the Drivers which refer to "the Host" + * doing things are inaccurate: the Launcher does all the device handling for + * the Guest. The Guest can't tell what's done by the the Launcher and what by + * the Host. + * + * Just to confuse you: to the Host kernel, the Launcher *is* the Guest and we + * shall see more of that later. + * + * We begin our understanding with the Host kernel interface which the Launcher + * uses: reading and writing a character device called /dev/lguest. All the + * work happens in the read(), write() and close() routines: */ static struct file_operations lguest_fops = { .owner = THIS_MODULE, .release = close, .write = write, .read = read, }; + +/* This is a textbook example of a "misc" character device. Populate a "struct + * miscdevice" and register it with misc_register(). */ static struct miscdevice lguest_dev = { .minor = MISC_DYNAMIC_MINOR, .name = "lguest",