lguest: documentation IV: Launcher
Documentation: The Launcher
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/drivers/lguest/core.c b/drivers/lguest/core.c
index 2cea0c8..1eb05f9 100644
--- a/drivers/lguest/core.c
+++ b/drivers/lguest/core.c
@@ -208,24 +208,39 @@
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 @@
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 @@
}
}
+/* 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 @@
|| 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 d2f02f0..da28812 100644
--- a/drivers/lguest/io.c
+++ b/drivers/lguest/io.c
@@ -27,8 +27,36 @@
#include <linux/uaccess.h>
#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 @@
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 @@
% 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 @@
&& 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 @@
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 @@
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 @@
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 @@
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 @@
}
}
+ /* 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 @@
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 @@
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 @@
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 @@
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 @@
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 @@
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 @@
up_read(fshared);
mutex_unlock(&lguest_lock);
}
+/*:*/
void release_all_dma(struct lguest *lg)
{
@@ -365,7 +553,8 @@
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 @@
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 @@
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 @@
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 3e2ddfb..3b9dc12 100644
--- a/drivers/lguest/lg.h
+++ b/drivers/lguest/lg.h
@@ -244,6 +244,30 @@
/* 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 @@
(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 6ae86f2..80d1b58 100644
--- a/drivers/lguest/lguest_user.c
+++ b/drivers/lguest/lguest_user.c
@@ -9,33 +9,62 @@
#include <linux/fs.h>
#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 @@
}
}
-/* + 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 @@
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 @@
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 @@
/* 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 @@
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 @@
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 @@
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 @@
}
}
+/*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",