[PATCH 16/16] per-vcpu lguest pgdir management

From: Glauber de Oliveira Costa
Date: Mon Jan 07 2008 - 08:11:46 EST


this patch makes the pgdir management per-vcpu. The pgdirs pool
is still guest-wide (although it'll probably need to grow when we
are really executing more vcpus), but the pgdidx index is gone,
since it makes no sense anymore. Instead, we use a per-vcpu
index.

Signed-off-by: Glauber de Oliveira Costa <gcosta@xxxxxxxxxx>
---
drivers/lguest/hypercalls.c | 2 +-
drivers/lguest/interrupts_and_traps.c | 6 ++--
drivers/lguest/lg.h | 12 +++---
drivers/lguest/page_tables.c | 60 +++++++++++++++++----------------
drivers/lguest/x86/core.c | 6 ++--
5 files changed, 44 insertions(+), 42 deletions(-)

diff --git a/drivers/lguest/hypercalls.c b/drivers/lguest/hypercalls.c
index ae8c0b4..b3a1942 100644
--- a/drivers/lguest/hypercalls.c
+++ b/drivers/lguest/hypercalls.c
@@ -60,7 +60,7 @@ static void do_hcall(struct lg_vcpu *vcpu, struct hcall_args *args)
if (args->arg1)
guest_pagetable_clear_all(vcpu);
else
- guest_pagetable_flush_user(lg);
+ guest_pagetable_flush_user(vcpu);
break;

/* All these calls simply pass the arguments through to the right
diff --git a/drivers/lguest/interrupts_and_traps.c b/drivers/lguest/interrupts_and_traps.c
index 745f3ae..68b403f 100644
--- a/drivers/lguest/interrupts_and_traps.c
+++ b/drivers/lguest/interrupts_and_traps.c
@@ -77,7 +77,7 @@ static void set_guest_interrupt(struct lg_vcpu *vcpu, u32 lo, u32 hi,
virtstack = vcpu->esp1;
ss = vcpu->ss1;

- origstack = gstack = guest_pa(lg, virtstack);
+ origstack = gstack = guest_pa(vcpu, virtstack);
/* We push the old stack segment and pointer onto the new
* stack: when the Guest does an "iret" back from the interrupt
* handler the CPU will notice they're dropping privilege
@@ -89,7 +89,7 @@ static void set_guest_interrupt(struct lg_vcpu *vcpu, u32 lo, u32 hi,
virtstack = vcpu->regs->esp;
ss = vcpu->regs->ss;

- origstack = gstack = guest_pa(lg, virtstack);
+ origstack = gstack = guest_pa(vcpu, virtstack);
}

/* Remember that we never let the Guest actually disable interrupts, so
@@ -325,7 +325,7 @@ void pin_stack_pages(struct lg_vcpu *vcpu)
* start of the page after the kernel stack. Subtract one to
* get back onto the first stack page, and keep subtracting to
* get to the rest of the stack pages. */
- pin_page(lg, vcpu->esp1 - 1 - i * PAGE_SIZE);
+ pin_page(vcpu, vcpu->esp1 - 1 - i * PAGE_SIZE);
}

/* Direct traps also mean that we need to know whenever the Guest wants to use
diff --git a/drivers/lguest/lg.h b/drivers/lguest/lg.h
index d33445f..6e6a69e 100644
--- a/drivers/lguest/lg.h
+++ b/drivers/lguest/lg.h
@@ -57,6 +57,8 @@ struct lg_vcpu {
unsigned long regs_page;
struct lguest_regs *regs;

+ int vcpu_pgd; /* which pgd this vcpu is currently using */
+
/* If a hypercall was asked for, this points to the arguments. */
struct hcall_args *hcall;
u32 next_hcall;
@@ -92,8 +94,6 @@ struct lguest
int changed;
struct lguest_pages *last_pages;

- /* We keep a small number of these. */
- u32 pgdidx;
struct pgdir pgdirs[4];

unsigned long noirq_start, noirq_end;
@@ -170,14 +170,14 @@ void free_guest_pagetable(struct lguest *lg);
void guest_new_pagetable(struct lg_vcpu *vcpu, unsigned long pgtable);
void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 i);
void guest_pagetable_clear_all(struct lg_vcpu *vcpu);
-void guest_pagetable_flush_user(struct lguest *lg);
+void guest_pagetable_flush_user(struct lg_vcpu *vcpu);
void guest_set_pte(struct lguest *lg, unsigned long gpgdir,
unsigned long vaddr, pte_t val);
void map_switcher_in_guest(struct lg_vcpu *vcpu,
struct lguest_pages *pages);
-int demand_page(struct lguest *info, unsigned long cr2, int errcode);
-void pin_page(struct lguest *lg, unsigned long vaddr);
-unsigned long guest_pa(struct lguest *lg, unsigned long vaddr);
+int demand_page(struct lg_vcpu *vcpu, unsigned long cr2, int errcode);
+void pin_page(struct lg_vcpu *vcpu, unsigned long vaddr);
+unsigned long guest_pa(struct lg_vcpu *vcpu, unsigned long vaddr);
void page_table_guest_data_init(struct lguest *lg);

/* <arch>/core.c: */
diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c
index 1a7ac3a..839ea27 100644
--- a/drivers/lguest/page_tables.c
+++ b/drivers/lguest/page_tables.c
@@ -94,10 +94,10 @@ static pte_t *spte_addr(struct lguest *lg, pgd_t spgd, unsigned long vaddr)

/* These two functions just like the above two, except they access the Guest
* page tables. Hence they return a Guest address. */
-static unsigned long gpgd_addr(struct lguest *lg, unsigned long vaddr)
+static unsigned long gpgd_addr(struct lg_vcpu *vcpu, unsigned long vaddr)
{
unsigned int index = vaddr >> (PGDIR_SHIFT);
- return lg->pgdirs[lg->pgdidx].gpgdir + index * sizeof(pgd_t);
+ return vcpu->lg->pgdirs[vcpu->vcpu_pgd].gpgdir + index * sizeof(pgd_t);
}

static unsigned long gpte_addr(struct lguest *lg,
@@ -200,22 +200,23 @@ static void check_gpgd(struct lguest *lg, pgd_t gpgd)
*
* If we fixed up the fault (ie. we mapped the address), this routine returns
* 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)
+int demand_page(struct lg_vcpu *vcpu, unsigned long vaddr, int errcode)
{
pgd_t gpgd;
pgd_t *spgd;
unsigned long gpte_ptr;
pte_t gpte;
pte_t *spte;
+ struct lguest *lg = vcpu->lg;

/* First step: get the top-level Guest page table entry. */
- gpgd = lgread(lg, gpgd_addr(lg, vaddr), pgd_t);
+ gpgd = lgread(lg, gpgd_addr(vcpu, vaddr), pgd_t);
/* Toplevel not present? We can't map it in. */
if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
return 0;

/* Now look at the matching shadow entry. */
- spgd = spgd_addr(lg, lg->pgdidx, vaddr);
+ spgd = spgd_addr(lg, vcpu->vcpu_pgd, vaddr);
if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) {
/* No shadow entry: allocate a new shadow PTE page. */
unsigned long ptepage = get_zeroed_page(GFP_KERNEL);
@@ -297,19 +298,19 @@ int demand_page(struct lguest *lg, unsigned long vaddr, int errcode)
*
* This is a quick version which answers the question: is this virtual address
* mapped by the shadow page tables, and is it writable? */
-static int page_writable(struct lguest *lg, unsigned long vaddr)
+static int page_writable(struct lg_vcpu *vcpu, unsigned long vaddr)
{
pgd_t *spgd;
unsigned long flags;

/* Look at the current top level entry: is it present? */
- spgd = spgd_addr(lg, lg->pgdidx, vaddr);
+ spgd = spgd_addr(vcpu->lg, vcpu->vcpu_pgd, vaddr);
if (!(pgd_flags(*spgd) & _PAGE_PRESENT))
return 0;

/* Check the flags on the pte entry itself: it must be present and
* writable. */
- flags = pte_flags(*(spte_addr(lg, *spgd, vaddr)));
+ flags = pte_flags(*(spte_addr(vcpu->lg, *spgd, vaddr)));

return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW);
}
@@ -317,10 +318,10 @@ static int page_writable(struct lguest *lg, unsigned long vaddr)
/* So, when pin_stack_pages() asks us to pin a page, we check if it's already
* in the page tables, and if not, we call demand_page() with error code 2
* (meaning "write"). */
-void pin_page(struct lguest *lg, unsigned long vaddr)
+void pin_page(struct lg_vcpu *vcpu, unsigned long vaddr)
{
- if (!page_writable(lg, vaddr) && !demand_page(lg, vaddr, 2))
- kill_guest(lg, "bad stack page %#lx", vaddr);
+ if (!page_writable(vcpu, vaddr) && !demand_page(vcpu, vaddr, 2))
+ kill_guest(vcpu->lg, "bad stack page %#lx", vaddr);
}

/*H:450 If we chase down the release_pgd() code, it looks like this: */
@@ -358,28 +359,28 @@ static void flush_user_mappings(struct lguest *lg, int idx)
*
* 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)
+void guest_pagetable_flush_user(struct lg_vcpu *vcpu)
{
/* Drop the userspace part of the current page table. */
- flush_user_mappings(lg, lg->pgdidx);
+ flush_user_mappings(vcpu->lg, vcpu->vcpu_pgd);
}
/*:*/

/* We walk down the guest page tables to get a guest-physical address */
-unsigned long guest_pa(struct lguest *lg, unsigned long vaddr)
+unsigned long guest_pa(struct lg_vcpu *vcpu, unsigned long vaddr)
{
pgd_t gpgd;
pte_t gpte;

/* First step: get the top-level Guest page table entry. */
- gpgd = lgread(lg, gpgd_addr(lg, vaddr), pgd_t);
+ gpgd = lgread(vcpu->lg, gpgd_addr(vcpu, vaddr), pgd_t);
/* Toplevel not present? We can't map it in. */
if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
- kill_guest(lg, "Bad address %#lx", vaddr);
+ kill_guest(vcpu->lg, "Bad address %#lx", vaddr);

- gpte = lgread(lg, gpte_addr(lg, gpgd, vaddr), pte_t);
+ gpte = lgread(vcpu->lg, gpte_addr(vcpu->lg, gpgd, vaddr), pte_t);
if (!(pte_flags(gpte) & _PAGE_PRESENT))
- kill_guest(lg, "Bad address %#lx", vaddr);
+ kill_guest(vcpu->lg, "Bad address %#lx", vaddr);

return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK);
}
@@ -399,11 +400,12 @@ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
/*H:435 And this is us, creating the new page directory. If we really do
* allocate a new one (and so the kernel parts are not there), we set
* blank_pgdir. */
-static unsigned int new_pgdir(struct lguest *lg,
+static unsigned int new_pgdir(struct lg_vcpu *vcpu,
unsigned long gpgdir,
int *blank_pgdir)
{
unsigned int next;
+ struct lguest *lg = vcpu->lg;

/* We pick one entry at random to throw out. Choosing the Least
* Recently Used might be better, but this is easy. */
@@ -413,7 +415,7 @@ static unsigned int new_pgdir(struct lguest *lg,
lg->pgdirs[next].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
/* If the allocation fails, just keep using the one we have */
if (!lg->pgdirs[next].pgdir)
- next = lg->pgdidx;
+ next = vcpu->vcpu_pgd;
else
/* This is a blank page, so there are no kernel
* mappings: caller must map the stack! */
@@ -442,9 +444,9 @@ void guest_new_pagetable(struct lg_vcpu *vcpu, unsigned long pgtable)
/* If not, we allocate or mug an existing one: if it's a fresh one,
* repin gets set to 1. */
if (newpgdir == ARRAY_SIZE(lg->pgdirs))
- newpgdir = new_pgdir(lg, pgtable, &repin);
+ newpgdir = new_pgdir(vcpu, pgtable, &repin);
/* Change the current pgd index to the new one. */
- lg->pgdidx = newpgdir;
+ vcpu->vcpu_pgd = newpgdir;
/* If it was completely blank, we map in the Guest kernel stack */
if (repin)
pin_stack_pages(vcpu);
@@ -591,11 +593,11 @@ int init_guest_pagetable(struct lguest *lg, unsigned long pgtable)
{
/* We start on the first shadow page table, and give it a blank PGD
* page. */
- lg->pgdidx = 0;
- lg->pgdirs[lg->pgdidx].gpgdir = pgtable;
- lg->pgdirs[lg->pgdidx].pgdir = (pgd_t*)get_zeroed_page(GFP_KERNEL);
- if (!lg->pgdirs[lg->pgdidx].pgdir)
+ lg->pgdirs[0].gpgdir = pgtable;
+ lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
+ if (!lg->pgdirs[0].pgdir)
return -ENOMEM;
+ lg->vcpus[0].vcpu_pgd = 0;
return 0;
}

@@ -607,7 +609,7 @@ void page_table_guest_data_init(struct lguest *lg)
/* We tell the Guest that it can't use the top 4MB of virtual
* addresses used by the Switcher. */
|| put_user(4U*1024*1024, &lg->lguest_data->reserve_mem)
- || put_user(lg->pgdirs[lg->pgdidx].gpgdir,&lg->lguest_data->pgdir))
+ || put_user(lg->pgdirs[0].gpgdir, &lg->lguest_data->pgdir))
kill_guest(lg, "bad guest page %p", lg->lguest_data);

/* In flush_user_mappings() we loop from 0 to
@@ -638,7 +640,6 @@ void free_guest_pagetable(struct lguest *lg)
void map_switcher_in_guest(struct lg_vcpu *vcpu,
struct lguest_pages *pages)
{
- struct lguest *lg = vcpu->lg;
pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages);
pgd_t switcher_pgd;
pte_t regs_pte;
@@ -648,7 +649,8 @@ void map_switcher_in_guest(struct lg_vcpu *vcpu,
* page for this CPU (with appropriate flags). */
switcher_pgd = __pgd(__pa(switcher_pte_page) | _PAGE_KERNEL);

- lg->pgdirs[lg->pgdidx].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
+ vcpu->lg->pgdirs[vcpu->vcpu_pgd].pgdir[SWITCHER_PGD_INDEX] =
+ switcher_pgd;

/* We also change the Switcher PTE page. When we're running the Guest,
* we want the Guest's "regs" page to appear where the first Switcher
diff --git a/drivers/lguest/x86/core.c b/drivers/lguest/x86/core.c
index 9a64174..b85e3de 100644
--- a/drivers/lguest/x86/core.c
+++ b/drivers/lguest/x86/core.c
@@ -147,7 +147,7 @@ static void run_guest_once(struct lg_vcpu *vcpu,
* 0-th argument above, ie "a"). %ebx contains the
* physical address of the Guest's top-level page
* directory. */
- : "0"(pages), "1"(__pa(lg->pgdirs[lg->pgdidx].pgdir))
+ : "0"(pages), "1"(__pa(lg->pgdirs[vcpu->vcpu_pgd].pgdir))
/* We tell gcc that all these registers could change,
* which means we don't have to save and restore them in
* the Switcher. */
@@ -225,7 +225,7 @@ static int emulate_insn(struct lg_vcpu *vcpu)
unsigned int insnlen = 0, in = 0, shift = 0;
/* The eip contains the *virtual* address of the Guest's instruction:
* guest_pa just subtracts the Guest's page_offset. */
- unsigned long physaddr = guest_pa(lg, vcpu->regs->eip);
+ unsigned long physaddr = guest_pa(vcpu, vcpu->regs->eip);

/* This must be the Guest kernel trying to do something, not userspace!
* The bottom two bits of the CS segment register are the privilege
@@ -307,7 +307,7 @@ void lguest_arch_handle_trap(struct lg_vcpu *vcpu)
*
* The errcode tells whether this was a read or a write, and
* whether kernel or userspace code. */
- if (demand_page(lg, vcpu->arch.last_pagefault,
+ if (demand_page(vcpu, vcpu->arch.last_pagefault,
vcpu->regs->errcode))
return;

--
1.5.0.6

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/