-reserved-ram for PCI passthrough without VT-d and without paravirt

Hello,

These three patches (one against host kernel, one against kvm.git, one against kvm-userland.git) forces KVM to map all RAM mapped in the virtualized e820 map provided to the guest with gfn = hfn. In turn it's now possible to give direct hardware access to the guest, all DMA will work fine on the virtualized guest ram.

The bios has to be updated to alter the end of the first ram slot in the virtualized e820 map. This is unfixable as the address hardcoded in the current bios is higher than what's marked as ram in my hardware e820 map.

The only exception where gfn != hfn for ranges included in the virtualized e820 map is for the magic bios page at host physical address zero (bytes from 0 to 4096). All linux versions will definitely never attempt to dma on such a page. If all OS are like linux there will be no problem and pci passthrough will work regardless of the guest OS without requiring any paravirtualization, nor VT-d.

This only implements the memory management side, the logic to map mmio regions into kvm address space will require further changes. The limit of the reserved ram is around 1G and it has to be set at compile time, so the guest will run with no more than 1G of ram (it's fairly easy to extend it to 2G though). You can't run more than one guest with -reserved-ram at once or they'll be overwriting themself. You need access to /dev/mem on the userland side, and CAP_ADMIN on the kernel side to run this.

I choosed an approach to require the minimal number of changes given this is a short term approach due the lack of hardware features in lots of cpus out there.

This is how the memory layout looks like when live guest runs (physical start set to 512M and kvm -m 300).

7f108d429000-7f108d42b000 rw-p 7f108d429000 00:00 0
7f108d42b000-7f108d4ba000 rw-s 00000000 00:0e 275 /dev/mem
7f108d4ba000-7f108d52a000 rw-p 7f108d4ba000 00:00 0
7f108d52a000-7f10a002a000 rw-s 00100000 00:0e 275 /dev/mem
7f10a002a000-7f10a142e000 rw-p 7f10a002a000 00:00 0

  PID TTY      STAT   TIME  MAJFL   TRS   DRS   RSS %MEM COMMAND
 5522 ?        SLl    4:06      1  1568 427067 33144  7.0 bin/x86_64/kvm/bin/qemu-system-x86_64 -hda tmp/vir

RSS isn't including the reserved ram of course.

-------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services for
just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace

This is the kvm.git patch to enable -reserved-ram (without this kvm will simply gracefully fail to emulate the illegal instruction inside the bad_page). This trick avoids altering the ioctl api with libkvm, in short if get_user_pages fails on a host kernel with reserved ram config option enabled, it tries to see if it's a remap_pfn_range mapping backing the memslot. In such case it checks if the ram was reserved with page_count 0 and if so it disables the reference counting as those pages are invisibile to linux. As long as pfn_valid returns, pfn_to_page should be safe, so shall the memmap array be allocated with holes corresponding to the holes generated in the e820 map, simply bad_page will be returned gracefully without risk like if this patch wasn't applied to kvm.git.

index e9ae5db..e7a9c82 100644
@@ -263,7 +263,8 @@ static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
 	npage = vcpu->arch.update_pte.page;
 	if (!npage)
 		return;
-	get_page(npage);
+	if (!page_is_reserved(npage))
+		get_page(npage);
 	mmu_set_spte(vcpu, spte, page->role.access, pte_access, 0, 0,
 		     gpte & PT_DIRTY_MASK, NULL, largepage, gpte_to_gfn(gpte),
 		     npage, true);

index f4e1436..7f087ac 100644
@@ -281,6 +281,18 @@ static inline void kvm_migrate_apic_timer(struct kvm_vcpu *vcpu)
 	set_bit(KVM_REQ_MIGRATE_TIMER, &vcpu->requests);
 }
 
+#ifdef CONFIG_RESERVE_PHYSICAL_START
+static inline int page_is_reserved(struct page * page)
+{
+	return !page_count(page);
+}
+#else /* CONFIG_RESERVE_PHYSICAL_START */
+static inline int page_is_reserved(struct page * page)
+{
+	return 0;
+}
+#endif /* CONFIG_RESERVE_PHYSICAL_START */
+
 enum kvm_stat_kind {
 	KVM_STAT_VM,
 	KVM_STAT_VCPU,

-------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services for
just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace

This is the kvm-userland.git patch overwriting the ranges in the virtualized e820 map with /dev/mem. All is validated through /proc/iomem, so shall the hardware e820 map be weird, there will be zero risk of corruption, simply it will fail to startup with a verbose error.

The bios has to be rebuilt to pass the variable address near 640k where to stop the virtualized e820 slot in function of the ram available in the host, and in function of the eary-reserve for things like the smp trampoline page that we don't want to pass as available ram to the guest. Only the page at address zero is magic and it's mapped as ram in the guest, but it's allocated through regular anonymous memory as you can see from the first /dev/mem mapping starting at area+reserved[0]. To rebuild the bios "make bios" before "make install" should do the trick. If you don't rebuild the bios everything will work fine if you don't use pci-passthrough, but then pci passthrough will randomly memory corrupt the host.

index 318de57..f93a6c6 100644
@@ -4251,6 +4251,7 @@ int15_function32(regs, ES, DS, FLAGS)
   Bit32u  extra_lowbits_memory_size=0;
   Bit16u  CX,DX;
   Bit8u   extra_highbits_memory_size=0;
+  Bit32u  below_640_end;
 
 BX_DEBUG_INT15("int15 AX=%04x\n",regs.u.r16.ax);
 

@@ -4305,6 +4306,11 @@ ASM_END
          case 0x20: // coded by osmaker aka K.J.
             if(regs.u.r32.edx == 0x534D4150)
             {
+                below_640_end = inb_cmos(0x16);
+                below_640_end <<= 8;
+                below_640_end |= inb_cmos(0x15);
+                below_640_end *= 1024;
+
                 extended_memory_size = inb_cmos(0x35);
                 extended_memory_size <<= 8;
                 extended_memory_size |= inb_cmos(0x34);

@@ -4334,7 +4340,7 @@ ASM_END
                 {
                     case 0:
                         set_e820_range(ES, regs.u.r16.di,
-                                       0x0000000L, 0x0009fc00L, 0, 0, 1);
+                                       0x0000000L, below_640_end, 0, 0, 1);
                         regs.u.r32.ebx = 1;
                         regs.u.r32.eax = 0x534D4150;
                         regs.u.r32.ecx = 0x14;

@@ -4343,7 +4349,7 @@ ASM_END
                         break;
                     case 1:
                         set_e820_range(ES, regs.u.r16.di,
-                                       0x0009fc00L, 0x000a0000L, 0, 0, 2);
+                                       below_640_end, 0x000a0000L, 0, 0, 2);
                         regs.u.r32.ebx = 2;
                         regs.u.r32.eax = 0x534D4150;
                         regs.u.r32.ecx = 0x14;

index 0d2e6c3..a6b28c8 100644
@@ -198,6 +198,8 @@ static void cmos_init(ram_addr_t ram_size, ram_addr_t above_4g_mem_size,
 
     /* memory size */
     val = 640; /* base memory in K */
+    if (reserved_ram)
+	    val = reserved[1] / 1024;
     rtc_set_memory(s, 0x15, val);
     rtc_set_memory(s, 0x16, val >> 8);
 

index 2e7d3e0..90d626d 100644
Binary files a/qemu/pc-bios/bios.bin and b/qemu/pc-bios/bios.bin differ

index c728605..db0dda4 100644
@@ -103,6 +103,8 @@ extern int autostart;
 extern int old_param;
 extern int hpagesize;
 extern const char *bootp_filename;
+extern int reserved_ram;
+extern int64_t reserved[4];
 
 
 #ifdef USE_KQEMU

-------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services for
just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace

Hello,

The "reserved RAM" can be mapped by virtualization software with /dev/mem to create a 1:1 mapping between guest physical (bus) address and host physical (bus) address. Please let me know if something like this can be merged in -mm (this is the minimal possible change to achieve the feature). The part at the end is more a fix, but it's only required with this applied (unless you want to have a kexec above ~40M).

Here the complete patchset:

http://marc.info/?l=kvm-devel&m=120698256716369&w=2
http://marc.info/?l=kvm-devel&m=120698299317253&w=2
http://marc.info/?l=kvm-devel&m=120698328617835&w=2

Note current mainline is buggy so this patch should also applied with -R for the host kernel to boot but hopefully the regression will be solved sooner than later (no reply yet though).

http://marc.info/?l=kvm-devel&m=120673375913890&w=2

andrea@svm ~ $ cat /proc/iomem | head
00000000-00000fff : reserved RAM failed
00001000-0008ffff : reserved RAM
00090000-00091fff : reserved RAM failed
00092000-0009efff : reserved RAM
0009f000-0009ffff : reserved
000cd600-000cffff : pnp 00:0d
000f0000-000fffff : reserved
00100000-1fffffff : reserved RAM
20000000-3dedffff : System RAM

@@ -1107,8 +1107,36 @@ config CRASH_DUMP
 	  (CONFIG_RELOCATABLE=y).
 	  For more details see Documentation/kdump/kdump.txt
 
+config RESERVE_PHYSICAL_START
+	bool "Reserve all RAM below PHYSICAL_START (EXPERIMENTAL)"
+	depends on !RELOCATABLE && X86_64
+	help
+	  This makes the kernel use only RAM above __PHYSICAL_START.
+	  All memory below __PHYSICAL_START will be left unused and
+	  marked as "reserved RAM" in /proc/iomem. The few special
+	  pages that can't be relocated at addresses above
+	  __PHYSICAL_START and that can't be guaranteed to be unused
+	  by the running kernel will be marked "reserved RAM failed"
+	  in /proc/iomem. Those may or may be not used by the kernel
+	  (for example SMP trampoline pages would only be used if
+	  CPU hotplug is enabled).
+
+	  The "reserved RAM" can be mapped by virtualization software
+	  with /dev/mem to create a 1:1 mapping between guest physical
+	  (bus) address and host physical (bus) address. This will
+	  allow PCI passthrough with DMA for the guest using the RAM
+	  with the 1:1 mapping. The only detail to take care of is the
+	  RAM marked "reserved RAM failed". The virtualization
+	  software must create for the guest an e820 map that only
+	  includes the "reserved RAM" regions but if the guest touches
+	  memory with guest physical address in the "reserved RAM
+	  failed" ranges (Linux guest will do that even if the RAM
+	  isn't present in the e820 map), it should provide that as
+	  RAM and map it with a non-linear mapping. This should allow
+	  any Linux kernel to run fine and hopefully any other OS too.
+
 config PHYSICAL_START
-	hex "Physical address where the kernel is loaded" if (EMBEDDED || CRASH_DUMP)
+	hex "Physical address where the kernel is loaded" if (EMBEDDED || CRASH_DUMP || RESERVE_PHYSICAL_START)
 	default "0x1000000" if X86_NUMAQ
 	default "0x200000" if X86_64
 	default "0x100000"

@@ -4,10 +4,19 @@
 #define E820MAX	128		/* number of entries in E820MAP */
 #define E820NR	0x1e8		/* # entries in E820MAP */
 
+#ifdef CONFIG_RESERVE_PHYSICAL_START
+#define E820_RESERVED_RAM 1
+#define E820_RESERVED_RAM_FAILED 2
+#define E820_RAM	3
+#define E820_RESERVED	4
+#define E820_ACPI	5
+#define E820_NVS	6
+#else
 #define E820_RAM	1
 #define E820_RESERVED	2
 #define E820_ACPI	3
 #define E820_NVS	4
+#endif
 
 #ifndef __ASSEMBLY__
 struct e820entry {

@@ -29,6 +29,7 @@
 #define __PAGE_OFFSET           _AC(0xffff810000000000, UL)
 
 #define __PHYSICAL_START	CONFIG_PHYSICAL_START
+#define __PHYSICAL_OFFSET	(__PHYSICAL_START-0x200000)
 #define __KERNEL_ALIGN		0x200000
 
 /*

@@ -51,7 +52,7 @@
  * Kernel image size is limited to 128 MB (see level2_kernel_pgt in
  * arch/x86/kernel/head_64.S), and it is mapped here:
  */
-#define KERNEL_IMAGE_SIZE	(128*1024*1024)
+#define KERNEL_IMAGE_SIZE	(128*1024*1024+__PHYSICAL_OFFSET)
 #define KERNEL_IMAGE_START	_AC(0xffffffff80000000, UL)
 
 #ifndef __ASSEMBLY__

@@ -140,7 +140,7 @@ static inline void native_pgd_clear(pgd_
 #define VMALLOC_START    _AC(0xffffc20000000000, UL)
 #define VMALLOC_END      _AC(0xffffe1ffffffffff, UL)
 #define VMEMMAP_START	 _AC(0xffffe20000000000, UL)
-#define MODULES_VADDR    _AC(0xffffffff88000000, UL)
+#define MODULES_VADDR    (0xffffffff88000000UL+__PHYSICAL_OFFSET)
 #define MODULES_END      _AC(0xfffffffffff00000, UL)
 #define MODULES_LEN   (MODULES_END - MODULES_VADDR)
 

@@ -47,7 +47,11 @@ static inline int cpu_present_to_apicid(
 
 #ifdef CONFIG_SMP
 
+#ifndef CONFIG_RESERVE_PHYSICAL_START
 #define SMP_TRAMPOLINE_BASE 0x6000
+#else
+#define SMP_TRAMPOLINE_BASE 0x90000 /* move it next to 640k */
+#endif
 
 extern int __cpu_disable(void);
 extern void __cpu_die(unsigned int cpu);

-------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services for
just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace

Hi Andrea,

Did you have to recompile the bios? How did you do that (or did you ask Avi to generate it?) Do you have a binary of the bios that I can use to test reserved ram?

Thanks,
Amit

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference Don't miss this year's exciting event. There's still time to save $100. Use priority code J8TL2D2.
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone

Yes.

make bios; make install should do the trick, the new bios should run after that.

thanks,
Andrea

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference Don't miss this year's exciting event. There's still time to save $100. Use priority code J8TL2D2.
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone

Well bcc, iasl on my machine have some problem.

Avi: do we have a machine where we have this setup for me to compile the bios?

Amit

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference Don't miss this year's exciting event. There's still time to save $100. Use priority code J8TL2D2.
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone

The versions I'm using are:

andrea@duo ~ $ esearch iasl dev86
[ Results for search key : iasl ]
[ Applications found : 1 ]

* sys-power/iasl
Latest version available: 20060912
Latest version installed: 20060912
Size of downloaded files: [no/bad digest]
Homepage: http://www.intel.com/technology/iapc/acpi/
Description: Intel ACPI Source Language (ASL) compiler
License: iASL

[ Results for search key : dev86 ]
[ Applications found : 1 ]

* sys-devel/dev86

      Latest version available: 0.16.17-r5
      Latest version installed: 0.16.17-r5
      Size of downloaded files: [no/bad digest]
      Homepage:    http://www.cix.co.uk/~mayday
      Description: Bruce's C compiler - Simple C compiler to generate
      8086 code
      License:     GPL-2

They worked fine.

I will send the patched bios to you by email in the meantime.

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference Don't miss this year's exciting event. There's still time to save $100. Use priority code J8TL2D2.
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone