}
/*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. */
+ * come wrapped up in the self-decompressing "bzImage" format. With a little
+ * work, we can load those, too. */
static unsigned long load_kernel(int fd)
{
Elf32_Ehdr hdr;
* to know where it is. */
return to_guest_phys(pgdir);
}
+/*:*/
/* Simple routine to roll all the commandline arguments together with spaces
* between them. */
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 base of guest "physical" memory, the top physical page to allow, the
+/*L:185 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 base of Guest "physical" memory, the top physical page to allow, the
* top level pagetable and the entry point for the Guest. */
static int tell_kernel(unsigned long pgdir, unsigned long start)
{
/*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.
+ * With console, block 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
+ * the LHREQ_BREAK command to the /dev/lguest file descriptor to tell the Host
+ * stop running the Guest. This causes the Launcher 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.
*
if (read(pipefd, &fd, sizeof(fd)) == 0)
exit(0);
/* Otherwise it's telling us to change what file
- * descriptors we're to listen to. */
+ * descriptors we're to listen to. Positive means
+ * listen to a new one, negative means stop
+ * listening. */
if (fd >= 0)
FD_SET(fd, &devices.infds);
else
{
int pipefd[2], child;
- /* We create a pipe to talk to the waker, and also so it knows when the
+ /* 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();
err(1, "forking");
if (child == 0) {
- /* Close the "writing" end of our copy of the pipe */
+ /* We are the Waker: close the "writing" end of our copy of the
+ * pipe and start waiting for input. */
close(pipefd[1]);
wake_parent(pipefd[0], lguest_fd);
}
return pipefd[1];
}
-/*L:210
+/*
* Device Handling.
*
- * When the Guest sends DMA to us, it sends us an array of addresses and sizes.
+ * When the Guest gives us a buffer, it sends 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
+ * we have a convenient routine which checks it and exits with an error message
* if something funny is going on:
*/
static void *_check_pointer(unsigned long addr, unsigned int size,
/* A macro which transparently hands the line number to the real function. */
#define check_pointer(addr,size) _check_pointer(addr, size, __LINE__)
-/* This function returns the next descriptor in the chain, or vq->vring.num. */
+/* Each buffer in the virtqueues is actually a chain of descriptors. This
+ * function returns the next descriptor in the chain, or vq->vring.num if we're
+ * at the end. */
static unsigned next_desc(struct virtqueue *vq, unsigned int i)
{
unsigned int next;
return head;
}
-/* Once we've used one of their buffers, we tell them about it. We'll then
+/* After we've used one of their buffers, we tell them about it. We'll then
* want to send them an interrupt, using trigger_irq(). */
static void add_used(struct virtqueue *vq, unsigned int head, int len)
{
struct vring_used_elem *used;
- /* Get a pointer to the next entry in the used ring. */
+ /* The virtqueue contains a ring of used buffers. Get a pointer to the
+ * next entry in that used ring. */
used = &vq->vring.used->ring[vq->vring.used->idx % vq->vring.num];
used->id = head;
used->len = len;
{
unsigned long buf[] = { LHREQ_IRQ, vq->config.irq };
+ /* If they don't want an interrupt, don't send one. */
if (vq->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
return;
trigger_irq(fd, vq);
}
-/* 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. */
+/*
+ * The Console
+ *
+ * Here is the input terminal setting we save, and the routine to restore them
+ * on exit so the user gets their terminal back. */
static struct termios orig_term;
static void restore_term(void)
{
}
}
-/* Handling output for network is also simple: we get all the output buffers
+/*
+ * The Network
+ *
+ * Handling output for network is also simple: we get all the output buffers
* and write them (ignoring the first element) to this device's file descriptor
* (stdout). */
static void handle_net_output(int fd, struct virtqueue *vq)
while ((head = get_vq_desc(vq, iov, &out, &in)) != vq->vring.num) {
if (in)
errx(1, "Input buffers in output queue?");
- /* Check header, but otherwise ignore it (we said we supported
- * no features). */
+ /* Check header, but otherwise ignore it (we told the Guest we
+ * supported no features, so it shouldn't have anything
+ * interesting). */
(void)convert(&iov[0], struct virtio_net_hdr);
len = writev(vq->dev->fd, iov+1, out-1);
add_used_and_trigger(fd, vq, head, len);
return true;
}
-/* This callback ensures we try again, in case we stopped console or net
+/*L:215 This is the callback attached to the network and console input
+ * virtqueues: it ensures we try again, in case we stopped console or net
* delivery because Guest didn't have any buffers. */
static void enable_fd(int fd, struct virtqueue *vq)
{
strnlen(from_guest_phys(addr), guest_limit - addr));
}
-/* This is called when the waker wakes us up: check for incoming file
+/* This is called when the Waker wakes us up: check for incoming file
* descriptors. */
static void handle_input(int fd)
{
}
/* Each device descriptor is followed by some configuration information.
- * The first byte is a "status" byte for the Guest to report what's happening.
- * After that are fields: u8 type, u8 len, [... len bytes...].
+ * Each configuration field looks like: u8 type, u8 len, [... len bytes...].
*
* This routine adds a new field to an existing device's descriptor. It only
* works for the last device, but that's OK because that's how we use it. */
/* Link virtqueue back to device. */
vq->dev = dev;
- /* Set up handler. */
+ /* Set the routine to call when the Guest does something to this
+ * virtqueue. */
vq->handle_output = handle_output;
+
+ /* Set the "Don't Notify Me" flag if we don't have a handler */
if (!handle_output)
vq->vring.used->flags = VRING_USED_F_NO_NOTIFY;
}
/* This routine does all the creation and setup of a new device, including
- * caling new_dev_desc() to allocate the descriptor and device memory. */
+ * calling new_dev_desc() to allocate the descriptor and device memory. */
static struct device *new_device(const char *name, u16 type, int fd,
bool (*handle_input)(int, struct device *))
{
/* 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. */
+ * is eth0, the first block device /dev/vda, etc. */
*devices.lastdev = dev;
dev->next = NULL;
devices.lastdev = &dev->next;
/* The console needs two virtqueues: the input then the output. When
* they put something the input queue, we make sure we're listening to
* stdin. When they put something in the output queue, we write it to
- * stdout. */
+ * stdout. */
add_virtqueue(dev, VIRTQUEUE_NUM, enable_fd);
add_virtqueue(dev, VIRTQUEUE_NUM, handle_console_output);
verbose("attached to bridge: %s\n", br_name);
}
-
-/*
- * Block device.
+/* Our block (disk) device should be really simple: the Guest asks for a block
+ * number and we read or write that position in the file. Unfortunately, that
+ * was amazingly slow: the Guest waits until the read is finished before
+ * running anything else, even if it could have been doing useful work.
*
- * Serving a block device is really easy: the Guest asks for a block number and
- * we read or write that position in the file.
- *
- * Unfortunately, this is amazingly slow: the Guest waits until the read is
- * finished before running anything else, even if it could be doing useful
- * work. We could use async I/O, except it's reputed to suck so hard that
- * characters actually go missing from your code when you try to use it.
+ * We could use async I/O, except it's reputed to suck so hard that characters
+ * actually go missing from your code when you try to use it.
*
* So we farm the I/O out to thread, and communicate with it via a pipe. */
-/* This hangs off device->priv, with the data. */
+/* This hangs off device->priv. */
struct vblk_info
{
/* The size of the file. */
* Launcher triggers interrupt to Guest. */
int done_fd;
};
+/*:*/
-/* This is the core of the I/O thread. It returns true if it did something. */
+/*L:210
+ * The Disk
+ *
+ * Remember that the block device is handled by a separate I/O thread. We head
+ * straight into the core of that thread here:
+ */
static bool service_io(struct device *dev)
{
struct vblk_info *vblk = dev->priv;
struct iovec iov[dev->vq->vring.num];
off64_t off;
+ /* See if there's a request waiting. If not, nothing to do. */
head = get_vq_desc(dev->vq, iov, &out_num, &in_num);
if (head == dev->vq->vring.num)
return false;
+ /* Every block request should contain at least one output buffer
+ * (detailing the location on disk and the type of request) and one
+ * input buffer (to hold the result). */
if (out_num == 0 || in_num == 0)
errx(1, "Bad virtblk cmd %u out=%u in=%u",
head, out_num, in_num);
in = convert(&iov[out_num+in_num-1], struct virtio_blk_inhdr);
off = out->sector * 512;
- /* This is how we implement barriers. Pretty poor, no? */
+ /* The block device implements "barriers", where the Guest indicates
+ * that it wants all previous writes to occur before this write. We
+ * don't have a way of asking our kernel to do a barrier, so we just
+ * synchronize all the data in the file. Pretty poor, no? */
if (out->type & VIRTIO_BLK_T_BARRIER)
fdatasync(vblk->fd);
+ /* In general the virtio block driver is allowed to try SCSI commands.
+ * It'd be nice if we supported eject, for example, but we don't. */
if (out->type & VIRTIO_BLK_T_SCSI_CMD) {
fprintf(stderr, "Scsi commands unsupported\n");
in->status = VIRTIO_BLK_S_UNSUPP;
/* When this read fails, it means Launcher died, so we follow. */
while (read(vblk->workpipe[0], &c, 1) == 1) {
- /* We acknowledge each request immediately, to reduce latency,
+ /* We acknowledge each request immediately to reduce latency,
* rather than waiting until we've done them all. I haven't
* measured to see if it makes any difference. */
while (service_io(dev))
return 0;
}
-/* When the thread says some I/O is done, we interrupt the Guest. */
+/* Now we've seen the I/O thread, we return to the Launcher to see what happens
+ * when the thread tells us it's completed some I/O. */
static bool handle_io_finish(int fd, struct device *dev)
{
char c;
- /* If child died, presumably it printed message. */
+ /* If the I/O thread died, presumably it printed the error, so we
+ * simply exit. */
if (read(dev->fd, &c, 1) != 1)
exit(1);
return true;
}
-/* When the Guest submits some I/O, we wake the I/O thread. */
+/* When the Guest submits some I/O, we just need to wake the I/O thread. */
static void handle_virtblk_output(int fd, struct virtqueue *vq)
{
struct vblk_info *vblk = vq->dev->priv;
exit(1);
}
-/* This creates a virtual block device. */
+/*L:198 This actually sets up a virtual block device. */
static void setup_block_file(const char *filename)
{
int p[2];
/* The device responds to return from I/O thread. */
dev = new_device("block", VIRTIO_ID_BLOCK, p[0], handle_io_finish);
- /* The device has a virtqueue. */
+ /* The device has one virtqueue, where the Guest places requests. */
add_virtqueue(dev, VIRTQUEUE_NUM, handle_virtblk_output);
/* Allocate the room for our own bookkeeping */
/* The I/O thread writes to this end of the pipe when done. */
vblk->done_fd = p[1];
- /* This is how we tell the I/O thread about more work. */
+ /* This is the second pipe, which is how we tell the I/O thread about
+ * more work. */
pipe(vblk->workpipe);
/* Create stack for thread and run it */
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.
+ /* 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. */
+ /* Service input, then unset the BREAK to release the Waker. */
handle_input(lguest_fd);
if (write(lguest_fd, args, sizeof(args)) < 0)
err(1, "Resetting break");
}
}
/*
- * This is the end of the Launcher.
+ * This is the end of the Launcher. The good news: we are over halfway
+ * through! The bad news: the most fiendish part of the code still lies ahead
+ * of us.
*
- * 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... :*/
+ * Are you ready? Take a deep breath and join me in the core of the Host, in
+ * "make Host".
+ :*/
static struct option opts[] = {
{ "verbose", 0, NULL, 'v' },
/* Memory, top-level pagetable, code startpoint and size of the
* (optional) initrd. */
unsigned long mem = 0, pgdir, start, initrd_size = 0;
- /* A temporary and the /dev/lguest file descriptor. */
+ /* Two temporaries and the /dev/lguest file descriptor. */
int i, c, lguest_fd;
/* The boot information for the Guest. */
struct boot_params *boot;
/* The boot header contains a command line pointer: we put the command
* line after the boot header. */
boot->hdr.cmd_line_ptr = to_guest_phys(boot + 1);
+ /* We use a simple helper to copy the arguments separated by spaces. */
concat((char *)(boot + 1), argv+optind+2);
/* Boot protocol version: 2.07 supports the fields for lguest. */
* When lazy_mode is set, it means we're allowed to defer all hypercalls and do
* them as a batch when lazy_mode is eventually turned off. Because hypercalls
* are reasonably expensive, batching them up makes sense. For example, a
- * large mmap might update dozens of page table entries: that code calls
+ * large munmap might update dozens of page table entries: that code calls
* paravirt_enter_lazy_mmu(), does the dozen updates, then calls
* lguest_leave_lazy_mode().
*
/*:*/
/*G:033
- * Here are our first native-instruction replacements: four functions for
- * interrupt control.
+ * After that diversion we return to our first native-instruction
+ * replacements: four functions for interrupt control.
*
* The simplest way of implementing these would be to have "turn interrupts
* off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow:
return lguest_data.irq_enabled;
}
-/* "restore_flags" just sets the flags back to the value given. */
+/* restore_flags() just sets the flags back to the value given. */
static void restore_fl(unsigned long flags)
{
lguest_data.irq_enabled = flags;
* it. The Host needs to know when the Guest wants to change them, so we have
* a whole series of functions like read_cr0() and write_cr0().
*
- * We start with CR0. CR0 allows you to turn on and off all kinds of basic
+ * We start with cr0. cr0 allows you to turn on and off all kinds of basic
* features, but Linux only really cares about one: the horrifically-named Task
* Switched (TS) bit at bit 3 (ie. 8)
*
current_cr0 &= ~X86_CR0_TS;
}
-/* CR2 is the virtual address of the last page fault, which the Guest only ever
+/* cr2 is the virtual address of the last page fault, which the Guest only ever
* reads. The Host kindly writes this into our "struct lguest_data", so we
* just read it out of there. */
static unsigned long lguest_read_cr2(void)
return lguest_data.cr2;
}
-/* CR3 is the current toplevel pagetable page: the principle is the same as
+/* cr3 is the current toplevel pagetable page: the principle is the same as
* cr0. Keep a local copy, and tell the Host when it changes. */
static void lguest_write_cr3(unsigned long cr3)
{
return current_cr3;
}
-/* CR4 is used to enable and disable PGE, but we don't care. */
+/* cr4 is used to enable and disable PGE, but we don't care. */
static unsigned long lguest_read_cr4(void)
{
return 0;
* maps virtual addresses to physical addresses using "page tables". We could
* use one huge index of 1 million entries: each address is 4 bytes, so that's
* 1024 pages just to hold the page tables. But since most virtual addresses
- * are unused, we use a two level index which saves space. The CR3 register
+ * are unused, we use a two level index which saves space. The cr3 register
* contains the physical address of the top level "page directory" page, which
* contains physical addresses of up to 1024 second-level pages. Each of these
* second level pages contains up to 1024 physical addresses of actual pages,
*
* Here's a diagram, where arrows indicate physical addresses:
*
- * CR3 ---> +---------+
+ * cr3 ---> +---------+
* | --------->+---------+
* | | | PADDR1 |
* Top-level | | PADDR2 |
*
* ... except in early boot when the kernel sets up the initial pagetables,
* which makes booting astonishingly slow. So we don't even tell the Host
- * anything changed until we've done the first page table switch.
- */
+ * anything changed until we've done the first page table switch. */
static void lguest_set_pte(pte_t *ptep, pte_t pteval)
{
*ptep = pteval;
/* Set up the timer interrupt (0) to go to our simple timer routine */
set_irq_handler(0, lguest_time_irq);
- /* Our clock structure look like arch/i386/kernel/tsc.c if we can use
- * the TSC, otherwise it's a dumb nanosecond-resolution clock. Either
- * way, the "rating" is initialized so high that it's always chosen
- * over any other clocksource. */
+ /* Our clock structure looks like arch/x86/kernel/tsc_32.c if we can
+ * use the TSC, otherwise it's a dumb nanosecond-resolution clock.
+ * Either way, the "rating" is set so high that it's always chosen over
+ * any other clocksource. */
if (lguest_data.tsc_khz)
lguest_clock.mult = clocksource_khz2mult(lguest_data.tsc_khz,
lguest_clock.shift);
* to work. They're pretty simple.
*/
-/* The Guest needs to tell the host what stack it expects traps to use. For
+/* The Guest needs to tell the Host what stack it expects traps to use. For
* native hardware, this is part of the Task State Segment mentioned above in
* lguest_load_tr_desc(), but to help hypervisors there's this special call.
*
return "LGUEST";
}
-/* Before virtqueues are set up, we use LHCALL_NOTIFY on normal memory to
- * produce console output. */
+/* We will eventually use the virtio console device to produce console output,
+ * but before that is set up we use LHCALL_NOTIFY on normal memory to produce
+ * console output. */
static __init int early_put_chars(u32 vtermno, const char *buf, int count)
{
char scratch[17];
unsigned int len = count;
+ /* We use a nul-terminated string, so we have to make a copy. Icky,
+ * huh? */
if (len > sizeof(scratch) - 1)
len = sizeof(scratch) - 1;
scratch[len] = '\0';
* Our current solution is to allow the paravirt back end to optionally patch
* over the indirect calls to replace them with something more efficient. We
* patch the four most commonly called functions: disable interrupts, enable
- * interrupts, restore interrupts and save interrupts. We usually have 10
+ * interrupts, restore interrupts and save interrupts. We usually have 6 or 10
* bytes to patch into: the Guest versions of these operations are small enough
* that we can fit comfortably.
*
asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory");
/* The Host uses the top of the Guest's virtual address space for the
- * Host<->Guest Switcher, and it tells us how much it needs in
+ * Host<->Guest Switcher, and it tells us how big that is in
* lguest_data.reserve_mem, set up on the LGUEST_INIT hypercall. */
reserve_top_address(lguest_data.reserve_mem);
/*
* This marks the end of stage II of our journey, The Guest.
*
- * It is now time for us to explore the nooks and crannies of the three Guest
- * devices and complete our understanding of the Guest in "make Drivers".
+ * It is now time for us to explore the layer of virtual drivers and complete
+ * our understanding of the Guest in "make Drivers".
*/
#include <asm/processor-flags.h>
/*G:020 This is where we begin: head.S notes that the boot header's platform
- * type field is "1" (lguest), so calls us here. The boot header is in %esi.
+ * type field is "1" (lguest), so calls us here.
*
* WARNING: be very careful here! We're running at addresses equal to physical
* addesses (around 0), not above PAGE_OFFSET as most code expectes
* boot. */
.section .init.text, "ax", @progbits
ENTRY(lguest_entry)
- /* Make initial hypercall now, so we can set up the pagetables. */
+ /* We make the "initialization" hypercall now to tell the Host about
+ * us, and also find out where it put our page tables. */
movl $LHCALL_LGUEST_INIT, %eax
movl $lguest_data - __PAGE_OFFSET, %edx
int $LGUEST_TRAP_ENTRY
/* The Host put the toplevel pagetable in lguest_data.pgdir. The movsl
- * instruction uses %esi implicitly. */
+ * instruction uses %esi implicitly as the source for the copy we'
+ * about to do. */
movl lguest_data - __PAGE_OFFSET + LGUEST_DATA_pgdir, %esi
/* Copy first 32 entries of page directory to __PAGE_OFFSET entries.
__free_pages(switcher_page[i], 0);
}
-/*L:305
+/*H:032
* Dealing With Guest Memory.
*
+ * Before we go too much further into the Host, we need to grok the routines
+ * we use to deal 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 the corresponding place in
* the memory region allocated by the Launcher.
lg->pending_notify = args->arg1;
break;
default:
+ /* It should be an architecture-specific hypercall. */
if (lguest_arch_do_hcall(lg, args))
kill_guest(lg, "Bad hypercall %li\n", args->arg0);
}
* Guest makes a hypercall, we end up here to set things up: */
static void initialize(struct lguest *lg)
{
-
/* You can't do anything until you're initialized. The Guest knows the
* rules, so we're unforgiving here. */
if (lg->hcall->arg0 != LHCALL_LGUEST_INIT) {
|| get_user(lg->noirq_end, &lg->lguest_data->noirq_end))
kill_guest(lg, "bad guest page %p", lg->lguest_data);
- /* We write the current time into the Guest's data page once now. */
+ /* We write the current time into the Guest's data page once so it can
+ * set its clock. */
write_timestamp(lg);
/* page_tables.c will also do some setup. */
/* This is the one case where the above accesses might have been the
* first write to a Guest page. This may have caused a copy-on-write
- * fault, but the Guest might be referring to the old (read-only)
- * page. */
+ * fault, but the old page might be (read-only) in the Guest
+ * pagetable. */
guest_pagetable_clear_all(lg);
}
* Normally it doesn't matter: the Guest will run again and
* update the trap number before we come back here.
*
- * However, if we are signalled or the Guest sends DMA to the
+ * However, if we are signalled or the Guest sends I/O to the
* Launcher, the run_guest() loop will exit without running the
* Guest. When it comes back it would try to re-run the
* hypercall. */
/* Remember that we never let the Guest actually disable interrupts, so
* the "Interrupt Flag" bit is always set. We copy that bit from the
- * Guest's "irq_enabled" field into the eflags word: the Guest copies
- * it back in "lguest_iret". */
+ * Guest's "irq_enabled" field into the eflags word: we saw the Guest
+ * copy it back in "lguest_iret". */
eflags = lg->regs->eflags;
if (get_user(irq_enable, &lg->lguest_data->irq_enabled) == 0
&& !(irq_enable & X86_EFLAGS_IF))
kill_guest(lg, "Disabling interrupts");
}
-/*H:200
+/*H:205
* Virtual Interrupts.
*
* maybe_do_interrupt() gets called before every entry to the Guest, to see if
* bogus one in): if we fail here, the Guest will be killed. */
if (!idt_present(lg->arch.idt[num].a, lg->arch.idt[num].b))
return 0;
- set_guest_interrupt(lg, lg->arch.idt[num].a, lg->arch.idt[num].b, has_err(num));
+ set_guest_interrupt(lg, lg->arch.idt[num].a, lg->arch.idt[num].b,
+ has_err(num));
return 1;
}
/*H:250 Here's the hard part: returning to the Host every time a trap happens
* and then calling deliver_trap() and re-entering the Guest is slow.
- * Particularly because Guest userspace system calls are traps (trap 128).
+ * Particularly because Guest userspace system calls are traps (usually trap
+ * 128).
*
* So we'd like to set up the IDT to tell the CPU to deliver traps directly
* into the Guest. This is possible, but the complexities cause the size of
* this file to double! However, 150 lines of code is worth writing for taking
* system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
- * the other hypervisors would tease it.
+ * the other hypervisors would beat it up at lunchtime.
*
* This routine indicates if a particular trap number could be delivered
* directly. */
* change stacks on each context switch. */
void guest_set_stack(struct lguest *lg, u32 seg, u32 esp, unsigned int pages)
{
- /* You are not allowd have a stack segment with privilege level 0: bad
+ /* You are not allowed have a stack segment with privilege level 0: bad
* Guest! */
if ((seg & 0x3) != GUEST_PL)
kill_guest(lg, "bad stack segment %i", seg);
* part of the Host: page table handling. */
/*H:235 This is the routine which actually checks the Guest's IDT entry and
- * transfers it into our entry in "struct lguest": */
+ * transfers it into the entry in "struct lguest": */
static void set_trap(struct lguest *lg, struct desc_struct *trap,
unsigned int num, u32 lo, u32 hi)
{
}
}
+/*H:200
+ * The Guest Clock.
+ *
+ * There are two sources of virtual interrupts. We saw one in lguest_user.c:
+ * the Launcher sending interrupts for virtual devices. The other is the Guest
+ * timer interrupt.
+ *
+ * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
+ * the next timer interrupt (in nanoseconds). We use the high-resolution timer
+ * infrastructure to set a callback at that time.
+ *
+ * 0 means "turn off the clock". */
void guest_set_clockevent(struct lguest *lg, unsigned long delta)
{
ktime_t expires;
return;
}
+ /* We use wallclock time here, so the Guest might not be running for
+ * all the time between now and the timer interrupt it asked for. This
+ * is almost always the right thing to do. */
expires = ktime_add_ns(ktime_get_real(), delta);
hrtimer_start(&lg->hrt, expires, HRTIMER_MODE_ABS);
}
+/* This is the function called when the Guest's timer expires. */
static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
{
struct lguest *lg = container_of(timer, struct lguest, hrt);
+ /* Remember the first interrupt is the timer interrupt. */
set_bit(0, lg->irqs_pending);
+ /* If the Guest is actually stopped, we need to wake it up. */
if (lg->halted)
wake_up_process(lg->tsk);
return HRTIMER_NORESTART;
}
+/* This sets up the timer for this Guest. */
void init_clockdev(struct lguest *lg)
{
hrtimer_init(&lg->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
void __lgread(struct lguest *, void *, unsigned long, unsigned);
void __lgwrite(struct lguest *, unsigned long, const void *, unsigned);
-/*L:306 Using memory-copy operations like that is usually inconvient, so we
+/*H:035 Using memory-copy operations like that is usually inconvient, so we
* have the following helper macros which read and write a specific type (often
* an unsigned long).
*
* 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
+ * There are many cases where the Guest can do 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
* Device configurations
*
* The configuration information for a device consists of a series of fields.
- * The device will look for these fields during setup.
+ * We don't really care what they are: the Launcher set them up, and the driver
+ * will look at them during setup.
*
* For us these fields come immediately after that device's descriptor in the
* lguest_devices page.
* The other piece of infrastructure virtio needs is a "virtqueue": a way of
* the Guest device registering buffers for the other side to read from or
* write into (ie. send and receive buffers). Each device can have multiple
- * virtqueues: for example the console has one queue for sending and one for
- * receiving.
+ * virtqueues: for example the console driver uses one queue for sending and
+ * another for receiving.
*
* Fortunately for us, a very fast shared-memory-plus-descriptors virtqueue
* already exists in virtio_ring.c. We just need to connect it up.
*
* This is kind of an ugly duckling. It'd be nicer to have a standard
* representation of a virtqueue in the configuration space, but it seems that
- * everyone wants to do it differently. The KVM guys want the Guest to
+ * everyone wants to do it differently. The KVM coders want the Guest to
* allocate its own pages and tell the Host where they are, but for lguest it's
* simpler for the Host to simply tell us where the pages are.
*
{
struct lguest_device *ldev;
+ /* Start with zeroed memory; Linux's device layer seems to count on
+ * it. */
ldev = kzalloc(sizeof(*ldev), GFP_KERNEL);
if (!ldev) {
printk(KERN_EMERG "Cannot allocate lguest dev %u\n",
#include <linux/fs.h>
#include "lg.h"
-/*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. */
+/*L:055 When something happens, the Waker process needs a way to stop the
+ * kernel running the Guest and return to the Launcher. So the Waker writes
+ * LHREQ_BREAK and the value "1" to /dev/lguest to do this. Once the Launcher
+ * has done whatever needs attention, it writes LHREQ_BREAK and "0" to release
+ * the Waker. */
static int break_guest_out(struct lguest *lg, const unsigned long __user *input)
{
unsigned long on;
- /* Fetch whether they're turning break on or off.. */
+ /* Fetch whether they're turning break on or off. */
if (get_user(on, input) != 0)
return -EFAULT;
if (on) {
lg->break_out = 1;
- /* Pop it out (may be running on different CPU) */
+ /* Pop it out of the Guest (may be running on different CPU) */
wake_up_process(lg->tsk);
/* Wait for them to reset it */
return wait_event_interruptible(lg->break_wq, !lg->break_out);
if (!lg)
return -EINVAL;
- /* If you're not the task which owns the guest, go away. */
+ /* If you're not the task which owns the Guest, go away. */
if (current != lg->tsk)
return -EPERM;
* base: The start of the Guest-physical memory inside the Launcher memory.
*
* 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.
+ * allowed to access. The Guest memory lives inside the Launcher, so it sets
+ * this to ensure the Guest can only reach its own memory.
*
* pgdir: The (Guest-physical) address of the top of the initial Guest
* pagetables (which are set up by the Launcher).
}
/*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
+ * start with an unsigned long 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 send interrupts. */
static ssize_t write(struct file *file, const char __user *in,
* 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.
+ * the Guest, but the Guest can't know that.
*
* Just to confuse you: to the Host kernel, the Launcher *is* the Guest and we
* shall see more of that later.
*
* We use two-level page tables for the Guest. If you're not entirely
* comfortable with virtual addresses, physical addresses and page tables then
- * I recommend you review lguest.c's "Page Table Handling" (with diagrams!).
+ * I recommend you review arch/x86/lguest/boot.c's "Page Table Handling" (with
+ * diagrams!).
*
* The Guest keeps page tables, but we maintain the actual ones here: these are
* called "shadow" page tables. Which is a very Guest-centric name: these are
*
* Anyway, this is the most complicated part of the Host code. There are seven
* parts to this:
- * (i) Setting up a page table entry for the Guest when it faults,
- * (ii) Setting up the page table entry for the Guest stack,
- * (iii) Setting up a page table entry when the Guest tells us it has changed,
+ * (i) Looking up a page table entry when the Guest faults,
+ * (ii) Making sure the Guest stack is mapped,
+ * (iii) Setting up a page table entry when the Guest tells us one has changed,
* (iv) Switching page tables,
- * (v) Flushing (thowing away) page tables,
+ * (v) Flushing (throwing away) page tables,
* (vi) Mapping the Switcher when the Guest is about to run,
* (vii) Setting up the page tables initially.
:*/
static DEFINE_PER_CPU(pte_t *, switcher_pte_pages);
#define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
-/*H:320 With our shadow and Guest types established, we need to deal with
- * them: the page table code is curly enough to need helper functions to keep
- * it clear and clean.
+/*H:320 The page table code is curly enough to need helper functions to keep it
+ * clear and clean.
*
* There are two functions which return pointers to the shadow (aka "real")
* page tables.
*
* spgd_addr() takes the virtual address and returns a pointer to the top-level
- * page directory entry for that address. Since we keep track of several page
- * tables, the "i" argument tells us which one we're interested in (it's
+ * page directory entry (PGD) for that address. Since we keep track of several
+ * page tables, the "i" argument tells us which one we're interested in (it's
* usually the current one). */
static pgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr)
{
return &lg->pgdirs[i].pgdir[index];
}
-/* This routine then takes the PGD entry given above, which contains the
- * address of the PTE page. It then returns a pointer to the PTE entry for the
- * given address. */
+/* This routine then takes the page directory entry returned above, which
+ * contains the address of the page table entry (PTE) page. It then returns a
+ * pointer to the PTE entry for the given address. */
static pte_t *spte_addr(struct lguest *lg, pgd_t spgd, unsigned long vaddr)
{
pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
}
/*H:330
- * (i) Setting up a page table entry for the Guest when it faults
+ * (i) Looking up a page table entry when the Guest faults.
*
* We saw this call in run_guest(): when we see a page fault in the Guest, we
* come here. That's because we only set up the shadow page tables lazily as
* and return to the Guest without it knowing.
*
* If we fixed up the fault (ie. we mapped the address), this routine returns
- * true. */
+ * true. Otherwise, it was a real fault and we need to tell the Guest. */
int demand_page(struct lguest *lg, unsigned long vaddr, int errcode)
{
pgd_t gpgd;
if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW))
return 0;
- /* User access to a kernel page? (bit 3 == user access) */
+ /* User access to a kernel-only page? (bit 3 == user access) */
if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER))
return 0;
/* Check that the Guest PTE flags are OK, and the page number is below
* the pfn_limit (ie. not mapping the Launcher binary). */
check_gpte(lg, gpte);
+
/* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
gpte = pte_mkyoung(gpte);
-
if (errcode & 2)
gpte = pte_mkdirty(gpte);
else
/* If this is a read, don't set the "writable" bit in the page
* table entry, even if the Guest says it's writable. That way
- * we come back here when a write does actually ocur, so we can
- * update the Guest's _PAGE_DIRTY flag. */
+ * we will come back here when a write does actually occur, so
+ * we can update the Guest's _PAGE_DIRTY flag. */
*spte = gpte_to_spte(lg, pte_wrprotect(gpte), 0);
/* Finally, we write the Guest PTE entry back: we've set the
* _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */
lgwrite(lg, gpte_ptr, pte_t, gpte);
- /* We succeeded in mapping the page! */
+ /* The fault is fixed, the page table is populated, the mapping
+ * manipulated, the result returned and the code complete. A small
+ * delay and a trace of alliteration are the only indications the Guest
+ * has that a page fault occurred at all. */
return 1;
}
-/*H:360 (ii) Setting up the page table entry for the Guest stack.
+/*H:360
+ * (ii) Making sure the Guest stack is mapped.
*
- * Remember pin_stack_pages() which makes sure the stack is mapped? It could
- * simply call demand_page(), but as we've seen that logic is quite long, and
- * usually the stack pages are already mapped anyway, so it's not required.
+ * Remember that direct traps into the Guest need a mapped Guest kernel stack.
+ * pin_stack_pages() calls us here: we could simply call demand_page(), but as
+ * we've seen that logic is quite long, and usually the stack pages are already
+ * mapped, so it's overkill.
*
* This is a quick version which answers the question: is this virtual address
* mapped by the shadow page tables, and is it writable? */
pgd_t *spgd;
unsigned long flags;
- /* Look at the top level entry: is it present? */
+ /* Look at the current top level entry: is it present? */
spgd = spgd_addr(lg, lg->pgdidx, vaddr);
if (!(pgd_flags(*spgd) & _PAGE_PRESENT))
return 0;
release_pte(ptepage[i]);
/* Now we can free the page of PTEs */
free_page((long)ptepage);
- /* And zero out the PGD entry we we never release it twice. */
+ /* And zero out the PGD entry so we never release it twice. */
*spgd = __pgd(0);
}
}
-/*H:440 (v) Flushing (thowing away) page tables,
- *
- * We saw flush_user_mappings() called when we re-used a top-level pgdir page.
- * It simply releases every PTE page from 0 up to the kernel address. */
+/*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings()
+ * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
+ * It simply releases every PTE page from 0 up to the Guest's kernel address. */
static void flush_user_mappings(struct lguest *lg, int idx)
{
unsigned int i;
release_pgd(lg, lg->pgdirs[idx].pgdir + i);
}
-/* The Guest also has a hypercall to do this manually: it's used when a large
- * number of mappings have been changed. */
+/*H:440 (v) Flushing (throwing away) page tables,
+ *
+ * The Guest has a hypercall to throw away the page tables: it's used when a
+ * large number of mappings have been changed. */
void guest_pagetable_flush_user(struct lguest *lg)
{
/* Drop the userspace part of the current page table. */
/*H:430 (iv) Switching page tables
*
- * This is what happens when the Guest changes page tables (ie. changes the
- * top-level pgdir). This happens on almost every context switch. */
+ * Now we've seen all the page table setting and manipulation, let's see what
+ * what happens when the Guest changes page tables (ie. changes the top-level
+ * pgdir). This occurs on almost every context switch. */
void guest_new_pagetable(struct lguest *lg, unsigned long pgtable)
{
int newpgdir, repin = 0;
}
/*H:470 Finally, a routine which throws away everything: all PGD entries in all
- * the shadow page tables. This is used when we destroy the Guest. */
+ * the shadow page tables, including the Guest's kernel mappings. This is used
+ * when we destroy the Guest. */
static void release_all_pagetables(struct lguest *lg)
{
unsigned int i, j;
/* We also throw away everything when a Guest tells us it's changed a kernel
* mapping. Since kernel mappings are in every page table, it's easiest to
- * throw them all away. This is amazingly slow, but thankfully rare. */
+ * throw them all away. This traps the Guest in amber for a while as
+ * everything faults back in, but it's rare. */
void guest_pagetable_clear_all(struct lguest *lg)
{
release_all_pagetables(lg);
/* We need the Guest kernel stack mapped again. */
pin_stack_pages(lg);
}
+/*:*/
+/*M:009 Since we throw away all mappings when a kernel mapping changes, our
+ * performance sucks for guests using highmem. In fact, a guest with
+ * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
+ * usually slower than a Guest with less memory.
+ *
+ * This, of course, cannot be fixed. It would take some kind of... well, I
+ * don't know, but the term "puissant code-fu" comes to mind. :*/
/*H:420 This is the routine which actually sets the page table entry for then
* "idx"'th shadow page table.
static void do_set_pte(struct lguest *lg, int idx,
unsigned long vaddr, pte_t gpte)
{
- /* Look up the matching shadow page directot entry. */
+ /* Look up the matching shadow page directory entry. */
pgd_t *spgd = spgd_addr(lg, idx, vaddr);
/* If the top level isn't present, there's no entry to update. */
*spte = gpte_to_spte(lg, gpte,
pte_flags(gpte) & _PAGE_DIRTY);
} else
- /* Otherwise we can demand_page() it in later. */
+ /* Otherwise kill it and we can demand_page() it in
+ * later. */
*spte = __pte(0);
}
}
}
/*H:400
- * (iii) Setting up a page table entry when the Guest tells us it has changed.
+ * (iii) Setting up a page table entry when the Guest tells us one has changed.
*
* Just like we did in interrupts_and_traps.c, it makes sense for us to deal
* with the other side of page tables while we're here: what happens when the
/*H:480 (vi) Mapping the Switcher when the Guest is about to run.
*
- * The Switcher and the two pages for this CPU need to be available to the
+ * The Switcher and the two pages for this CPU need to be visible in the
* Guest (and not the pages for other CPUs). We have the appropriate PTE pages
- * for each CPU already set up, we just need to hook them in. */
+ * for each CPU already set up, we just need to hook them in now we know which
+ * Guest is about to run on this CPU. */
void map_switcher_in_guest(struct lguest *lg, struct lguest_pages *pages)
{
pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages);
__pgprot(_PAGE_PRESENT|_PAGE_ACCESSED));
}
+/* We've made it through the page table code. Perhaps our tired brains are
+ * still processing the details, or perhaps we're simply glad it's over.
+ *
+ * If nothing else, note that all this complexity in juggling shadow page
+ * tables in sync with the Guest's page tables is for one reason: for most
+ * Guests this page table dance determines how bad performance will be. This
+ * is why Xen uses exotic direct Guest pagetable manipulation, and why both
+ * Intel and AMD have implemented shadow page table support directly into
+ * hardware.
+ *
+ * There is just one file remaining in the Host. */
+
/*H:510 At boot or module load time, init_pagetables() allocates and populates
* the Switcher PTE page for each CPU. */
__init int init_pagetables(struct page **switcher_page, unsigned int pages)
#include "lg.h"
/*H:600
- * We've almost completed the Host; there's just one file to go!
- *
* Segments & The Global Descriptor Table
*
* (That title sounds like a bad Nerdcore group. Not to suggest that there are
|| num == GDT_ENTRY_DOUBLEFAULT_TSS);
}
-/*H:610 Once the GDT has been changed, we fix the new entries up a little. We
+/*H:630 Once the Guest gave us new GDT entries, we fix them up a little. We
* don't care if they're invalid: the worst that can happen is a General
* Protection Fault in the Switcher when it restores a Guest segment register
* which tries to use that entry. Then we kill the Guest for causing such a
}
}
-/* This routine is called at boot or modprobe time for each CPU to set up the
- * "constant" GDT entries for Guests running on that CPU. */
+/*H:610 Like the IDT, we never simply use the GDT the Guest gives us. We keep
+ * a GDT for each CPU, and copy across the Guest's entries each time we want to
+ * run the Guest on that CPU.
+ *
+ * This routine is called at boot or modprobe time for each CPU to set up the
+ * constant GDT entries: the ones which are the same no matter what Guest we're
+ * running. */
void setup_default_gdt_entries(struct lguest_ro_state *state)
{
struct desc_struct *gdt = state->guest_gdt;
unsigned long tss = (unsigned long)&state->guest_tss;
- /* The hypervisor segments are full 0-4G segments, privilege level 0 */
+ /* The Switcher segments are full 0-4G segments, privilege level 0 */
gdt[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT;
gdt[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT;
- /* The TSS segment refers to the TSS entry for this CPU, so we cannot
- * copy it from the Guest. Forgive the magic flags */
+ /* The TSS segment refers to the TSS entry for this particular CPU.
+ * Forgive the magic flags: the 0x8900 means the entry is Present, it's
+ * privilege level 0 Available 386 TSS system segment, and the 0x67
+ * means Saturn is eclipsed by Mercury in the twelfth house. */
gdt[GDT_ENTRY_TSS].a = 0x00000067 | (tss << 16);
gdt[GDT_ENTRY_TSS].b = 0x00008900 | (tss & 0xFF000000)
| ((tss >> 16) & 0x000000FF);
}
-/* This routine is called before the Guest is run for the first time. */
+/* This routine sets up the initial Guest GDT for booting. All entries start
+ * as 0 (unusable). */
void setup_guest_gdt(struct lguest *lg)
{
/* Start with full 0-4G segments... */
lg->arch.gdt[GDT_ENTRY_KERNEL_DS].b |= (GUEST_PL << 13);
}
-/* Like the IDT, we never simply use the GDT the Guest gives us. We set up the
- * GDTs for each CPU, then we copy across the entries each time we want to run
- * a different Guest on that CPU. */
-
-/* A partial GDT load, for the three "thead-local storage" entries. Otherwise
- * it's just like load_guest_gdt(). So much, in fact, it would probably be
- * neater to have a single hypercall to cover both. */
+/*H:650 An optimization of copy_gdt(), for just the three "thead-local storage"
+ * entries. */
void copy_gdt_tls(const struct lguest *lg, struct desc_struct *gdt)
{
unsigned int i;
gdt[i] = lg->arch.gdt[i];
}
-/* This is the full version */
+/*H:640 When the Guest is run on a different CPU, or the GDT entries have
+ * changed, copy_gdt() is called to copy the Guest's GDT entries across to this
+ * CPU's GDT. */
void copy_gdt(const struct lguest *lg, struct desc_struct *gdt)
{
unsigned int i;
gdt[i] = lg->arch.gdt[i];
}
-/* This is where the Guest asks us to load a new GDT (LHCALL_LOAD_GDT). */
+/*H:620 This is where the Guest asks us to load a new GDT (LHCALL_LOAD_GDT).
+ * We copy it from the Guest and tweak the entries. */
void load_guest_gdt(struct lguest *lg, unsigned long table, u32 num)
{
/* We assume the Guest has the same number of GDT entries as the
lg->changed |= CHANGED_GDT;
}
+/* This is the fast-track version for just changing the three TLS entries.
+ * Remember that this happens on every context switch, so it's worth
+ * optimizing. But wouldn't it be neater to have a single hypercall to cover
+ * both cases? */
void guest_load_tls(struct lguest *lg, unsigned long gtls)
{
struct desc_struct *tls = &lg->arch.gdt[GDT_ENTRY_TLS_MIN];
__lgread(lg, tls, gtls, sizeof(*tls)*GDT_ENTRY_TLS_ENTRIES);
fixup_gdt_table(lg, GDT_ENTRY_TLS_MIN, GDT_ENTRY_TLS_MAX+1);
+ /* Note that just the TLS entries have changed. */
lg->changed |= CHANGED_GDT_TLS;
}
+/*:*/
-/*
+/*H:660
* With this, we have finished the Host.
*
* Five of the seven parts of our task are complete. You have made it through
static DEFINE_PER_CPU(struct lguest *, last_guest);
/*S:010
- * We are getting close to the Switcher.
+ * We approach the Switcher.
*
* Remember that each CPU has two pages which are visible to the Guest when it
* runs on that CPU. This has to contain the state for that Guest: we copy the
*
* The lcall also pushes the old code segment (KERNEL_CS) onto the
* stack, then the address of this call. This stack layout happens to
- * exactly match the stack of an interrupt... */
+ * exactly match the stack layout created by an interrupt... */
asm volatile("pushf; lcall *lguest_entry"
/* This is how we tell GCC that %eax ("a") and %ebx ("b")
* are changed by this routine. The "=" means output. */
}
/*:*/
+/*M:002 There are hooks in the scheduler which we can register to tell when we
+ * get kicked off the CPU (preempt_notifier_register()). This would allow us
+ * to lazily disable SYSENTER which would regain some performance, and should
+ * also simplify copy_in_guest_info(). Note that we'd still need to restore
+ * things when we exit to Launcher userspace, but that's fairly easy.
+ *
+ * The hooks were designed for KVM, but we can also put them to good use. :*/
+
/*H:040 This is the i386-specific code to setup and run the Guest. Interrupts
* are disabled: we own the CPU. */
void lguest_arch_run_guest(struct lguest *lg)
{
- /* Remember the awfully-named TS bit? If the Guest has asked
- * to set it we set it now, so we can trap and pass that trap
- * to the Guest if it uses the FPU. */
+ /* Remember the awfully-named TS bit? If the Guest has asked to set it
+ * we set it now, so we can trap and pass that trap to the Guest if it
+ * uses the FPU. */
if (lg->ts)
lguest_set_ts();
- /* SYSENTER is an optimized way of doing system calls. We
- * can't allow it because it always jumps to privilege level 0.
- * A normal Guest won't try it because we don't advertise it in
- * CPUID, but a malicious Guest (or malicious Guest userspace
- * program) could, so we tell the CPU to disable it before
- * running the Guest. */
+ /* SYSENTER is an optimized way of doing system calls. We can't allow
+ * it because it always jumps to privilege level 0. A normal Guest
+ * won't try it because we don't advertise it in CPUID, but a malicious
+ * Guest (or malicious Guest userspace program) could, so we tell the
+ * CPU to disable it before running the Guest. */
if (boot_cpu_has(X86_FEATURE_SEP))
wrmsr(MSR_IA32_SYSENTER_CS, 0, 0);
- /* Now we actually run the Guest. It will pop back out when
- * something interesting happens, and we can examine its
- * registers to see what it was doing. */
+ /* Now we actually run the Guest. It will return when something
+ * interesting happens, and we can examine its registers to see what it
+ * was doing. */
run_guest_once(lg, lguest_pages(raw_smp_processor_id()));
- /* The "regs" pointer contains two extra entries which are not
- * really registers: a trap number which says what interrupt or
- * trap made the switcher code come back, and an error code
- * which some traps set. */
+ /* Note that the "regs" pointer contains two extra entries which are
+ * not really registers: a trap number which says what interrupt or
+ * trap made the switcher code come back, and an error code which some
+ * traps set. */
- /* If the Guest page faulted, then the cr2 register will tell
- * us the bad virtual address. We have to grab this now,
- * because once we re-enable interrupts an interrupt could
- * fault and thus overwrite cr2, or we could even move off to a
- * different CPU. */
+ /* If the Guest page faulted, then the cr2 register will tell us the
+ * bad virtual address. We have to grab this now, because once we
+ * re-enable interrupts an interrupt could fault and thus overwrite
+ * cr2, or we could even move off to a different CPU. */
if (lg->regs->trapnum == 14)
lg->arch.last_pagefault = read_cr2();
/* Similarly, if we took a trap because the Guest used the FPU,
wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0);
}
-/*H:130 Our Guest is usually so well behaved; it never tries to do things it
- * isn't allowed to. Unfortunately, Linux's paravirtual infrastructure isn't
- * quite complete, because it doesn't contain replacements for the Intel I/O
- * instructions. As a result, the Guest sometimes fumbles across one during
- * the boot process as it probes for various things which are usually attached
- * to a PC.
+/*H:130 Now we've examined the hypercall code; our Guest can make requests.
+ * Our Guest is usually so well behaved; it never tries to do things it isn't
+ * allowed to, and uses hypercalls instead. Unfortunately, Linux's paravirtual
+ * infrastructure isn't quite complete, because it doesn't contain replacements
+ * for the Intel I/O instructions. As a result, the Guest sometimes fumbles
+ * across one during the boot process as it probes for various things which are
+ * usually attached to a PC.
*
- * When the Guest uses one of these instructions, we get trap #13 (General
+ * When the Guest uses one of these instructions, we get a trap (General
* Protection Fault) and come here. We see if it's one of those troublesome
* instructions and skip over it. We return true if we did. */
static int emulate_insn(struct lguest *lg)
void lguest_arch_handle_trap(struct lguest *lg)
{
switch (lg->regs->trapnum) {
- case 13: /* We've intercepted a GPF. */
- /* Check if this was one of those annoying IN or OUT
- * instructions which we need to emulate. If so, we
- * just go back into the Guest after we've done it. */
+ case 13: /* We've intercepted a General Protection Fault. */
+ /* Check if this was one of those annoying IN or OUT
+ * instructions which we need to emulate. If so, we just go
+ * back into the Guest after we've done it. */
if (lg->regs->errcode == 0) {
if (emulate_insn(lg))
return;
}
break;
- case 14: /* We've intercepted a page fault. */
- /* The Guest accessed a virtual address that wasn't
- * mapped. This happens a lot: we don't actually set
- * up most of the page tables for the Guest at all when
- * we start: as it runs it asks for more and more, and
- * we set them up as required. In this case, we don't
- * even tell the Guest that the fault happened.
- *
- * The errcode tells whether this was a read or a
- * write, and whether kernel or userspace code. */
+ case 14: /* We've intercepted a Page Fault. */
+ /* The Guest accessed a virtual address that wasn't mapped.
+ * This happens a lot: we don't actually set up most of the
+ * page tables for the Guest at all when we start: as it runs
+ * it asks for more and more, and we set them up as
+ * required. In this case, we don't even tell the Guest that
+ * the fault happened.
+ *
+ * The errcode tells whether this was a read or a write, and
+ * whether kernel or userspace code. */
if (demand_page(lg, lg->arch.last_pagefault, lg->regs->errcode))
return;
- /* OK, it's really not there (or not OK): the Guest
- * needs to know. We write out the cr2 value so it
- * knows where the fault occurred.
- *
- * Note that if the Guest were really messed up, this
- * could happen before it's done the INITIALIZE
- * hypercall, so lg->lguest_data will be NULL */
+ /* OK, it's really not there (or not OK): the Guest needs to
+ * know. We write out the cr2 value so it knows where the
+ * fault occurred.
+ *
+ * Note that if the Guest were really messed up, this could
+ * happen before it's done the LHCALL_LGUEST_INIT hypercall, so
+ * lg->lguest_data could be NULL */
if (lg->lguest_data &&
put_user(lg->arch.last_pagefault, &lg->lguest_data->cr2))
kill_guest(lg, "Writing cr2");
break;
case 7: /* We've intercepted a Device Not Available fault. */
- /* If the Guest doesn't want to know, we already
- * restored the Floating Point Unit, so we just
- * continue without telling it. */
+ /* If the Guest doesn't want to know, we already restored the
+ * Floating Point Unit, so we just continue without telling
+ * it. */
if (!lg->ts)
return;
break;
return 0;
}
-/* Now we've examined the hypercall code; our Guest can make requests. There
- * is one other way we can do things for the Guest, as we see in
- * emulate_insn(). :*/
/*L:030 lguest_arch_setup_regs()
*
/* %esi points to our boot information, at physical address 0, so don't
* touch it. */
+
/* There are a couple of GDT entries the Guest expects when first
* booting. */
-
setup_guest_gdt(lg);
}
* are feeling invigorated and refreshed then the next, more challenging stage
* can be found in "make Guest". :*/
+/*M:012 Lguest is meant to be simple: my rule of thumb is that 1% more LOC must
+ * gain at least 1% more performance. Since neither LOC nor performance can be
+ * measured beforehand, it generally means implementing a feature then deciding
+ * if it's worth it. And once it's implemented, who can say no?
+ *
+ * This is why I haven't implemented this idea myself. I want to, but I
+ * haven't. You could, though.
+ *
+ * The main place where lguest performance sucks is Guest page faulting. When
+ * a Guest userspace process hits an unmapped page we switch back to the Host,
+ * walk the page tables, find it's not mapped, switch back to the Guest page
+ * fault handler, which calls a hypercall to set the page table entry, then
+ * finally returns to userspace. That's two round-trips.
+ *
+ * If we had a small walker in the Switcher, we could quickly check the Guest
+ * page table and if the page isn't mapped, immediately reflect the fault back
+ * into the Guest. This means the Switcher would have to know the top of the
+ * Guest page table and the page fault handler address.
+ *
+ * For simplicity, the Guest should only handle the case where the privilege
+ * level of the fault is 3 and probably only not present or write faults. It
+ * should also detect recursive faults, and hand the original fault to the
+ * Host (which is actually really easy).
+ *
+ * Two questions remain. Would the performance gain outweigh the complexity?
+ * And who would write the verse documenting it? :*/
+
+/*M:011 Lguest64 handles NMI. This gave me NMI envy (until I looked at their
+ * code). It's worth doing though, since it would let us use oprofile in the
+ * Host when a Guest is running. :*/
+
/*S:100
* Welcome to the Switcher itself!
*
// All saved and there's now five steps before us:
// Stack, GDT, IDT, TSS
- // And last of all the page tables are flipped.
+ // Then last of all the page tables are flipped.
// Yet beware that our stack pointer must be
// Always valid lest an NMI hits
lgdt LGUEST_PAGES_guest_gdt_desc(%eax)
// The Guest's IDT we did partially
- // Move to the "struct lguest_pages" as well.
+ // Copy to "struct lguest_pages" as well.
lidt LGUEST_PAGES_guest_idt_desc(%eax)
// The TSS entry which controls traps
// Must be loaded up with "ltr" now:
+ // The GDT entry that TSS uses
+ // Changes type when we load it: damn Intel!
// For after we switch over our page tables
- // It (as the rest) will be writable no more.
- // (The GDT entry TSS needs
- // Changes type when we load it: damn Intel!)
+ // That entry will be read-only: we'd crash.
movl $(GDT_ENTRY_TSS*8), %edx
ltr %dx
// Look back now, before we take this last step!
// The Host's TSS entry was also marked used;
- // Let's clear it again, ere we return.
+ // Let's clear it again for our return.
// The GDT descriptor of the Host
// Points to the table after two "size" bytes
movl (LGUEST_PAGES_host_gdt_desc+2)(%eax), %edx
- // Clear the type field of "used" (byte 5, bit 2)
+ // Clear "used" from type field (byte 5, bit 2)
andb $0xFD, (GDT_ENTRY_TSS*8 + 5)(%edx)
// Once our page table's switched, the Guest is live!
// The page table change did one tricky thing:
// The Guest's register page has been mapped
- // Writable onto our %esp (stack) --
+ // Writable under our %esp (stack) --
// We can simply pop off all Guest regs.
popl %eax
popl %ebx
addl $8, %esp
// The last five stack slots hold return address
- // And everything needed to change privilege
- // Into the Guest privilege level of 1,
+ // And everything needed to switch privilege
+ // From Switcher's level 0 to Guest's 1,
// And the stack where the Guest had last left it.
// Interrupts are turned back on: we are Guest.
iret
-// There are two paths where we switch to the Host
+// We treat two paths to switch back to the Host
+// Yet both must save Guest state and restore Host
// So we put the routine in a macro.
-// We are on our way home, back to the Host
-// Interrupted out of the Guest, we come here.
#define SWITCH_TO_HOST \
/* We save the Guest state: all registers first \
* Laid out just as "struct lguest_regs" defines */ \
movl %esp, %eax; \
andl $(~(1 << PAGE_SHIFT - 1)), %eax; \
/* Save our trap number: the switch will obscure it \
- * (The Guest regs are not mapped here in the Host) \
+ * (In the Host the Guest regs are not mapped here) \
* %ebx holds it safe for deliver_to_host */ \
movl LGUEST_PAGES_regs_trapnum(%eax), %ebx; \
/* The Host GDT, IDT and stack! \
/* Switch to Host's GDT, IDT. */ \
lgdt LGUEST_PAGES_host_gdt_desc(%eax); \
lidt LGUEST_PAGES_host_idt_desc(%eax); \
- /* Restore the Host's stack where it's saved regs lie */ \
+ /* Restore the Host's stack where its saved regs lie */ \
movl LGUEST_PAGES_host_sp(%eax), %esp; \
- /* Last the TSS: our Host is complete */ \
+ /* Last the TSS: our Host is returned */ \
movl $(GDT_ENTRY_TSS*8), %edx; \
ltr %dx; \
/* Restore now the regs saved right at the first. */ \
popl %ds; \
popl %es
-// Here's where we come when the Guest has just trapped:
-// (Which trap we'll see has been pushed on the stack).
+// The first path is trod when the Guest has trapped:
+// (Which trap it was has been pushed on the stack).
// We need only switch back, and the Host will decode
// Why we came home, and what needs to be done.
return_to_host:
SWITCH_TO_HOST
iret
+// We are lead to the second path like so:
// An interrupt, with some cause external
// Has ajerked us rudely from the Guest's code
// Again we must return home to the Host
// But now we must go home via that place
// Where that interrupt was supposed to go
// Had we not been ensconced, running the Guest.
- // Here we see the cleverness of our stack:
+ // Here we see the trickness of run_guest_once():
// The Host stack is formed like an interrupt
// With EIP, CS and EFLAGS layered.
// Interrupt handlers end with "iret"
xorw %ax, %ax
orl %eax, %edx
// Now the address of the handler's in %edx
- // We call it now: its "iret" takes us home.
+ // We call it now: its "iret" drops us home.
jmp *%edx
// Every interrupt can come to us here
#define LHCALL_LOAD_TLS 16
#define LHCALL_NOTIFY 17
+#define LGUEST_TRAP_ENTRY 0x1F
+
+#ifndef __ASSEMBLY__
+#include <asm/hw_irq.h>
+
/*G:031 First, how does our Guest contact the Host to ask for privileged
* operations? There are two ways: the direct way is to make a "hypercall",
* to make requests of the Host Itself.
*
* Our hypercall mechanism uses the highest unused trap code (traps 32 and
- * above are used by real hardware interrupts). Seventeen hypercalls are
+ * above are used by real hardware interrupts). Fifteen hypercalls are
* available: the hypercall number is put in the %eax register, and the
* arguments (when required) are placed in %edx, %ebx and %ecx. If a return
* value makes sense, it's returned in %eax.
* Grossly invalid calls result in Sudden Death at the hands of the vengeful
* Host, rather than returning failure. This reflects Winston Churchill's
* definition of a gentleman: "someone who is only rude intentionally". */
-#define LGUEST_TRAP_ENTRY 0x1F
-
-#ifndef __ASSEMBLY__
-#include <asm/hw_irq.h>
-
static inline unsigned long
hcall(unsigned long call,
unsigned long arg1, unsigned long arg2, unsigned long arg3)
{
/* "int" is the Intel instruction to trigger a trap. */
asm volatile("int $" __stringify(LGUEST_TRAP_ENTRY)
- /* The call is in %eax (aka "a"), and can be replaced */
+ /* The call in %eax (aka "a") might be overwritten */
: "=a"(call)
- /* The other arguments are in %eax, %edx, %ebx & %ecx */
+ /* The arguments are in %eax, %edx, %ebx & %ecx */
: "a"(call), "d"(arg1), "b"(arg2), "c"(arg3)
/* "memory" means this might write somewhere in memory.
* This isn't true for all calls, but it's safe to tell
#define LG_CLOCK_MAX_DELTA ULONG_MAX
/*G:032 The second method of communicating with the Host is to via "struct
- * lguest_data". The Guest's very first hypercall is to tell the Host where
- * this is, and then the Guest and Host both publish information in it. :*/
+ * lguest_data". Once the Guest's initialization hypercall tells the Host where
+ * this is, the Guest and Host both publish information in it. :*/
struct lguest_data
{
/* 512 == enabled (same as eflags in normal hardware). The Guest
* real devices (think of the damage it could do!) we provide virtual devices.
* We could emulate a PCI bus with various devices on it, but that is a fairly
* complex burden for the Host and suboptimal for the Guest, so we have our own
- * "lguest" bus and simple drivers.
+ * simple lguest bus and we use "virtio" drivers. These drivers need a set of
+ * routines from us which will actually do the virtual I/O, but they handle all
+ * the net/block/console stuff themselves. This means that if we want to add
+ * a new device, we simply need to write a new virtio driver and create support
+ * for it in the Launcher: this code won't need to change.
*
* Devices are described by a simplified ID, a status byte, and some "config"
* bytes which describe this device's configuration. This is placed by the
projects / GitHub / mt8127 / android_kernel_alcatel_ttab.git / commitdiff