From a91d74a3c4de8115295ee87350c13a329164aaaf Mon Sep 17 00:00:00 2001 From: Rusty Russell Date: Thu, 30 Jul 2009 16:03:45 -0600 Subject: lguest: update commentry Every so often, after code shuffles, I need to go through and unbitrot the Lguest Journey (see drivers/lguest/README). Since we now use RCU in a simple form in one place I took the opportunity to expand that explanation. Signed-off-by: Rusty Russell Cc: Ingo Molnar Cc: Paul McKenney --- drivers/lguest/core.c | 7 ++- drivers/lguest/hypercalls.c | 6 ++- drivers/lguest/lguest_device.c | 11 +++-- drivers/lguest/lguest_user.c | 100 +++++++++++++++++++++++++++++++++++---- drivers/lguest/page_tables.c | 84 ++++++++++++++++++++++++-------- drivers/lguest/x86/core.c | 2 +- drivers/lguest/x86/switcher_32.S | 6 +-- 7 files changed, 176 insertions(+), 40 deletions(-) (limited to 'drivers/lguest') diff --git a/drivers/lguest/core.c b/drivers/lguest/core.c index cd058bc903ff..1e2cb846b3c9 100644 --- a/drivers/lguest/core.c +++ b/drivers/lguest/core.c @@ -217,10 +217,15 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user) /* * It's possible the Guest did a NOTIFY hypercall to the - * Launcher, in which case we return from the read() now. + * Launcher. */ if (cpu->pending_notify) { + /* + * Does it just needs to write to a registered + * eventfd (ie. the appropriate virtqueue thread)? + */ if (!send_notify_to_eventfd(cpu)) { + /* OK, we tell the main Laucher. */ if (put_user(cpu->pending_notify, user)) return -EFAULT; return sizeof(cpu->pending_notify); diff --git a/drivers/lguest/hypercalls.c b/drivers/lguest/hypercalls.c index 787ab4bc09f0..83511eb0923d 100644 --- a/drivers/lguest/hypercalls.c +++ b/drivers/lguest/hypercalls.c @@ -59,7 +59,7 @@ static void do_hcall(struct lg_cpu *cpu, struct hcall_args *args) case LHCALL_SHUTDOWN: { char msg[128]; /* - * Shutdown is such a trivial hypercall that we do it in four + * Shutdown is such a trivial hypercall that we do it in five * lines right here. * * If the lgread fails, it will call kill_guest() itself; the @@ -245,6 +245,10 @@ static void initialize(struct lg_cpu *cpu) * device), the Guest will still see the old page. In practice, this never * happens: why would the Guest read a page which it has never written to? But * a similar scenario might one day bite us, so it's worth mentioning. + * + * Note that if we used a shared anonymous mapping in the Launcher instead of + * mapping /dev/zero private, we wouldn't worry about cop-on-write. And we + * need that to switch the Launcher to processes (away from threads) anyway. :*/ /*H:100 diff --git a/drivers/lguest/lguest_device.c b/drivers/lguest/lguest_device.c index cc000e79c3d1..1401c1ace1ec 100644 --- a/drivers/lguest/lguest_device.c +++ b/drivers/lguest/lguest_device.c @@ -236,7 +236,7 @@ static void lg_notify(struct virtqueue *vq) extern void lguest_setup_irq(unsigned int irq); /* - * This routine finds the first virtqueue described in the configuration of + * This routine finds the Nth virtqueue described in the configuration of * this device and sets it up. * * This is kind of an ugly duckling. It'd be nicer to have a standard @@ -244,9 +244,6 @@ extern void lguest_setup_irq(unsigned int irq); * 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. - * - * So we provide drivers with a "find the Nth virtqueue and set it up" - * function. */ static struct virtqueue *lg_find_vq(struct virtio_device *vdev, unsigned index, @@ -422,7 +419,11 @@ static void add_lguest_device(struct lguest_device_desc *d, /* This devices' parent is the lguest/ dir. */ ldev->vdev.dev.parent = lguest_root; - /* We have a unique device index thanks to the dev_index counter. */ + /* + * The device type comes straight from the descriptor. There's also a + * device vendor field in the virtio_device struct, which we leave as + * 0. + */ ldev->vdev.id.device = d->type; /* * We have a simple set of routines for querying the device's diff --git a/drivers/lguest/lguest_user.c b/drivers/lguest/lguest_user.c index 7e92017103dc..b4d3f7ca554f 100644 --- a/drivers/lguest/lguest_user.c +++ b/drivers/lguest/lguest_user.c @@ -1,9 +1,8 @@ -/*P:200 - * This contains all the /dev/lguest code, whereby the userspace launcher +/*P:200 This contains all the /dev/lguest code, whereby the userspace launcher * controls and communicates with the Guest. For example, the first write will - * tell us the Guest's memory layout, pagetable, entry point and kernel address - * offset. A read will run the Guest until something happens, such as a signal - * or the Guest doing a NOTIFY out to the Launcher. + * tell us the Guest's memory layout and entry point. A read will run the + * Guest until something happens, such as a signal or the Guest doing a NOTIFY + * out to the Launcher. :*/ #include #include @@ -13,14 +12,41 @@ #include #include "lg.h" +/*L:056 + * Before we move on, let's jump ahead and look at what the kernel does when + * it needs to look up the eventfds. That will complete our picture of how we + * use RCU. + * + * The notification value is in cpu->pending_notify: we return true if it went + * to an eventfd. + */ bool send_notify_to_eventfd(struct lg_cpu *cpu) { unsigned int i; struct lg_eventfd_map *map; - /* lg->eventfds is RCU-protected */ + /* + * This "rcu_read_lock()" helps track when someone is still looking at + * the (RCU-using) eventfds array. It's not actually a lock at all; + * indeed it's a noop in many configurations. (You didn't expect me to + * explain all the RCU secrets here, did you?) + */ rcu_read_lock(); + /* + * rcu_dereference is the counter-side of rcu_assign_pointer(); it + * makes sure we don't access the memory pointed to by + * cpu->lg->eventfds before cpu->lg->eventfds is set. Sounds crazy, + * but Alpha allows this! Paul McKenney points out that a really + * aggressive compiler could have the same effect: + * http://lists.ozlabs.org/pipermail/lguest/2009-July/001560.html + * + * So play safe, use rcu_dereference to get the rcu-protected pointer: + */ map = rcu_dereference(cpu->lg->eventfds); + /* + * Simple array search: even if they add an eventfd while we do this, + * we'll continue to use the old array and just won't see the new one. + */ for (i = 0; i < map->num; i++) { if (map->map[i].addr == cpu->pending_notify) { eventfd_signal(map->map[i].event, 1); @@ -28,14 +54,43 @@ bool send_notify_to_eventfd(struct lg_cpu *cpu) break; } } + /* We're done with the rcu-protected variable cpu->lg->eventfds. */ rcu_read_unlock(); + + /* If we cleared the notification, it's because we found a match. */ return cpu->pending_notify == 0; } +/*L:055 + * One of the more tricksy tricks in the Linux Kernel is a technique called + * Read Copy Update. Since one point of lguest is to teach lguest journeyers + * about kernel coding, I use it here. (In case you're curious, other purposes + * include learning about virtualization and instilling a deep appreciation for + * simplicity and puppies). + * + * We keep a simple array which maps LHCALL_NOTIFY values to eventfds, but we + * add new eventfds without ever blocking readers from accessing the array. + * The current Launcher only does this during boot, so that never happens. But + * Read Copy Update is cool, and adding a lock risks damaging even more puppies + * than this code does. + * + * We allocate a brand new one-larger array, copy the old one and add our new + * element. Then we make the lg eventfd pointer point to the new array. + * That's the easy part: now we need to free the old one, but we need to make + * sure no slow CPU somewhere is still looking at it. That's what + * synchronize_rcu does for us: waits until every CPU has indicated that it has + * moved on to know it's no longer using the old one. + * + * If that's unclear, see http://en.wikipedia.org/wiki/Read-copy-update. + */ static int add_eventfd(struct lguest *lg, unsigned long addr, int fd) { struct lg_eventfd_map *new, *old = lg->eventfds; + /* + * We don't allow notifications on value 0 anyway (pending_notify of + * 0 means "nothing pending"). + */ if (!addr) return -EINVAL; @@ -62,12 +117,20 @@ static int add_eventfd(struct lguest *lg, unsigned long addr, int fd) } new->num++; - /* Now put new one in place. */ + /* + * Now put new one in place: rcu_assign_pointer() is a fancy way of + * doing "lg->eventfds = new", but it uses memory barriers to make + * absolutely sure that the contents of "new" written above is nailed + * down before we actually do the assignment. + * + * We have to think about these kinds of things when we're operating on + * live data without locks. + */ rcu_assign_pointer(lg->eventfds, new); /* * We're not in a big hurry. Wait until noone's looking at old - * version, then delete it. + * version, then free it. */ synchronize_rcu(); kfree(old); @@ -75,6 +138,14 @@ static int add_eventfd(struct lguest *lg, unsigned long addr, int fd) return 0; } +/*L:052 + * Receiving notifications from the Guest is usually done by attaching a + * particular LHCALL_NOTIFY value to an event filedescriptor. The eventfd will + * become readable when the Guest does an LHCALL_NOTIFY with that value. + * + * This is really convenient for processing each virtqueue in a separate + * thread. + */ static int attach_eventfd(struct lguest *lg, const unsigned long __user *input) { unsigned long addr, fd; @@ -86,6 +157,11 @@ static int attach_eventfd(struct lguest *lg, const unsigned long __user *input) if (get_user(fd, input) != 0) return -EFAULT; + /* + * Just make sure two callers don't add eventfds at once. We really + * only need to lock against callers adding to the same Guest, so using + * the Big Lguest Lock is overkill. But this is setup, not a fast path. + */ mutex_lock(&lguest_lock); err = add_eventfd(lg, addr, fd); mutex_unlock(&lguest_lock); @@ -106,6 +182,10 @@ static int user_send_irq(struct lg_cpu *cpu, const unsigned long __user *input) if (irq >= LGUEST_IRQS) return -EINVAL; + /* + * Next time the Guest runs, the core code will see if it can deliver + * this interrupt. + */ set_interrupt(cpu, irq); return 0; } @@ -307,10 +387,10 @@ unlock: * The first operation the Launcher does must be a write. All writes * 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. + * writes of other values to send interrupts or set up receipt of notifications. * * Note that we overload the "offset" in the /dev/lguest file to indicate what - * CPU number we're dealing with. Currently this is always 0, since we only + * CPU number we're dealing with. Currently this is always 0 since we only * support uniprocessor Guests, but you can see the beginnings of SMP support * here. */ diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c index 3da902e4b4cb..a8d0aee3bc0e 100644 --- a/drivers/lguest/page_tables.c +++ b/drivers/lguest/page_tables.c @@ -29,10 +29,10 @@ /*H:300 * The Page Table Code * - * 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 arch/x86/lguest/boot.c's "Page Table Handling" (with - * diagrams!). + * We use two-level page tables for the Guest, or three-level with PAE. If + * you're not entirely comfortable with virtual addresses, physical addresses + * and page tables then 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 @@ -52,9 +52,8 @@ :*/ /* - * 1024 entries in a page table page maps 1024 pages: 4MB. The Switcher is - * conveniently placed at the top 4MB, so it uses a separate, complete PTE - * page. + * The Switcher uses the complete top PTE page. That's 1024 PTE entries (4MB) + * or 512 PTE entries with PAE (2MB). */ #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1) @@ -81,7 +80,8 @@ static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); /*H:320 * The page table code is curly enough to need helper functions to keep it - * clear and clean. + * clear and clean. The kernel itself provides many of them; one advantage + * of insisting that the Guest and Host use the same CONFIG_PAE setting. * * There are two functions which return pointers to the shadow (aka "real") * page tables. @@ -155,7 +155,7 @@ static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) } /* - * These two functions just like the above two, except they access the Guest + * These functions are just like the above two, except they access the Guest * page tables. Hence they return a Guest address. */ static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) @@ -165,6 +165,7 @@ static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) } #ifdef CONFIG_X86_PAE +/* Follow the PGD to the PMD. */ static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr) { unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; @@ -172,6 +173,7 @@ static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr) return gpage + pmd_index(vaddr) * sizeof(pmd_t); } +/* Follow the PMD to the PTE. */ static unsigned long gpte_addr(struct lg_cpu *cpu, pmd_t gpmd, unsigned long vaddr) { @@ -181,6 +183,7 @@ static unsigned long gpte_addr(struct lg_cpu *cpu, return gpage + pte_index(vaddr) * sizeof(pte_t); } #else +/* Follow the PGD to the PTE (no mid-level for !PAE). */ static unsigned long gpte_addr(struct lg_cpu *cpu, pgd_t gpgd, unsigned long vaddr) { @@ -314,6 +317,7 @@ bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) pte_t gpte; pte_t *spte; + /* Mid level for PAE. */ #ifdef CONFIG_X86_PAE pmd_t *spmd; pmd_t gpmd; @@ -391,6 +395,8 @@ bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) */ gpte_ptr = gpte_addr(cpu, gpgd, vaddr); #endif + + /* Read the actual PTE value. */ gpte = lgread(cpu, gpte_ptr, pte_t); /* If this page isn't in the Guest page tables, we can't page it in. */ @@ -507,6 +513,7 @@ void pin_page(struct lg_cpu *cpu, unsigned long vaddr) if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2)) kill_guest(cpu, "bad stack page %#lx", vaddr); } +/*:*/ #ifdef CONFIG_X86_PAE static void release_pmd(pmd_t *spmd) @@ -543,7 +550,11 @@ static void release_pgd(pgd_t *spgd) } #else /* !CONFIG_X86_PAE */ -/*H:450 If we chase down the release_pgd() code, it looks like this: */ +/*H:450 + * If we chase down the release_pgd() code, the non-PAE version looks like + * this. The PAE version is almost identical, but instead of calling + * release_pte it calls release_pmd(), which looks much like this. + */ static void release_pgd(pgd_t *spgd) { /* If the entry's not present, there's nothing to release. */ @@ -898,17 +909,21 @@ void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx) /* ... throw it away. */ release_pgd(lg->pgdirs[pgdir].pgdir + idx); } + #ifdef CONFIG_X86_PAE +/* For setting a mid-level, we just throw everything away. It's easy. */ void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx) { guest_pagetable_clear_all(&lg->cpus[0]); } #endif -/* - * Once we know how much memory we have we can construct simple identity (which +/*H:505 + * To get through boot, we construct simple identity page mappings (which * set virtual == physical) and linear mappings which will get the Guest far - * enough into the boot to create its own. + * enough into the boot to create its own. The linear mapping means we + * simplify the Guest boot, but it makes assumptions about their PAGE_OFFSET, + * as you'll see. * * We lay them out of the way, just below the initrd (which is why we need to * know its size here). @@ -944,6 +959,10 @@ static unsigned long setup_pagetables(struct lguest *lg, linear = (void *)pgdir - linear_pages * PAGE_SIZE; #ifdef CONFIG_X86_PAE + /* + * And the single mid page goes below that. We only use one, but + * that's enough to map 1G, which definitely gets us through boot. + */ pmds = (void *)linear - PAGE_SIZE; #endif /* @@ -957,13 +976,14 @@ static unsigned long setup_pagetables(struct lguest *lg, return -EFAULT; } +#ifdef CONFIG_X86_PAE /* - * The top level points to the linear page table pages above. - * We setup the identity and linear mappings here. + * Make the Guest PMD entries point to the corresponding place in the + * linear mapping (up to one page worth of PMD). */ -#ifdef CONFIG_X86_PAE for (i = j = 0; i < mapped_pages && j < PTRS_PER_PMD; i += PTRS_PER_PTE, j++) { + /* FIXME: native_set_pmd is overkill here. */ native_set_pmd(&pmd, __pmd(((unsigned long)(linear + i) - mem_base) | _PAGE_PRESENT | _PAGE_RW | _PAGE_USER)); @@ -971,18 +991,36 @@ static unsigned long setup_pagetables(struct lguest *lg, return -EFAULT; } + /* One PGD entry, pointing to that PMD page. */ set_pgd(&pgd, __pgd(((u32)pmds - mem_base) | _PAGE_PRESENT)); + /* Copy it in as the first PGD entry (ie. addresses 0-1G). */ if (copy_to_user(&pgdir[0], &pgd, sizeof(pgd)) != 0) return -EFAULT; + /* + * And the third PGD entry (ie. addresses 3G-4G). + * + * FIXME: This assumes that PAGE_OFFSET for the Guest is 0xC0000000. + */ if (copy_to_user(&pgdir[3], &pgd, sizeof(pgd)) != 0) return -EFAULT; #else + /* + * The top level points to the linear page table pages above. + * We setup the identity and linear mappings here. + */ phys_linear = (unsigned long)linear - mem_base; for (i = 0; i < mapped_pages; i += PTRS_PER_PTE) { pgd_t pgd; + /* + * Create a PGD entry which points to the right part of the + * linear PTE pages. + */ pgd = __pgd((phys_linear + i * sizeof(pte_t)) | (_PAGE_PRESENT | _PAGE_RW | _PAGE_USER)); + /* + * Copy it into the PGD page at 0 and PAGE_OFFSET. + */ if (copy_to_user(&pgdir[i / PTRS_PER_PTE], &pgd, sizeof(pgd)) || copy_to_user(&pgdir[pgd_index(PAGE_OFFSET) + i / PTRS_PER_PTE], @@ -992,8 +1030,8 @@ static unsigned long setup_pagetables(struct lguest *lg, #endif /* - * We return the top level (guest-physical) address: remember where - * this is. + * We return the top level (guest-physical) address: we remember where + * this is to write it into lguest_data when the Guest initializes. */ return (unsigned long)pgdir - mem_base; } @@ -1031,7 +1069,9 @@ int init_guest_pagetable(struct lguest *lg) lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL); if (!lg->pgdirs[0].pgdir) return -ENOMEM; + #ifdef CONFIG_X86_PAE + /* For PAE, we also create the initial mid-level. */ pgd = lg->pgdirs[0].pgdir; pmd_table = (pmd_t *) get_zeroed_page(GFP_KERNEL); if (!pmd_table) @@ -1040,11 +1080,13 @@ int init_guest_pagetable(struct lguest *lg) set_pgd(pgd + SWITCHER_PGD_INDEX, __pgd(__pa(pmd_table) | _PAGE_PRESENT)); #endif + + /* This is the current page table. */ lg->cpus[0].cpu_pgd = 0; return 0; } -/* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ +/*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ void page_table_guest_data_init(struct lg_cpu *cpu) { /* We get the kernel address: above this is all kernel memory. */ @@ -1105,12 +1147,16 @@ void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages) pmd_t switcher_pmd; pmd_t *pmd_table; + /* FIXME: native_set_pmd is overkill here. */ native_set_pmd(&switcher_pmd, pfn_pmd(__pa(switcher_pte_page) >> PAGE_SHIFT, PAGE_KERNEL_EXEC)); + /* Figure out where the pmd page is, by reading the PGD, and converting + * it to a virtual address. */ pmd_table = __va(pgd_pfn(cpu->lg-> pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX]) << PAGE_SHIFT); + /* Now write it into the shadow page table. */ native_set_pmd(&pmd_table[SWITCHER_PMD_INDEX], switcher_pmd); #else pgd_t switcher_pgd; diff --git a/drivers/lguest/x86/core.c b/drivers/lguest/x86/core.c index 96f7d88ec7f8..6ae388849a3b 100644 --- a/drivers/lguest/x86/core.c +++ b/drivers/lguest/x86/core.c @@ -187,7 +187,7 @@ static void run_guest_once(struct lg_cpu *cpu, struct lguest_pages *pages) * 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. * - * We could also try using this hooks for PGE, but that might be too expensive. + * We could also try using these hooks for PGE, but that might be too expensive. * * The hooks were designed for KVM, but we can also put them to good use. :*/ diff --git a/drivers/lguest/x86/switcher_32.S b/drivers/lguest/x86/switcher_32.S index 6dec09793836..40634b0db9f7 100644 --- a/drivers/lguest/x86/switcher_32.S +++ b/drivers/lguest/x86/switcher_32.S @@ -1,7 +1,7 @@ /*P:900 - * This is the Switcher: code which sits at 0xFFC00000 astride both the - * Host and Guest to do the low-level Guest<->Host switch. It is as simple as - * it can be made, but it's naturally very specific to x86. + * This is the Switcher: code which sits at 0xFFC00000 (or 0xFFE00000) astride + * both the Host and Guest to do the low-level Guest<->Host switch. It is as + * simple as it can be made, but it's naturally very specific to x86. * * You have now completed Preparation. If this has whet your appetite; if you * are feeling invigorated and refreshed then the next, more challenging stage -- cgit v1.2.3