Re: Re: Linux Stubdom Problem

[Keir Fraser <keir.xen@gmail.com>]

On 22/08/2011 20:36, "Stefano Stabellini" <stefano.stabellini@eu.citrix.com>
wrote:

> On Mon, 22 Aug 2011, Jiageng Yu wrote:
>> Hi Stefano,
>> 
>>      I am trying to fix the vram mapping issue. That is my new patch.
>> It is still needed to debug. Please check it for me and make sure I am
>> running on the right way.
>> 
>>     I define a new enmu type Stubdom_Vga_State which is used to notify
>> xen_remap_bucket whether to map the vram memory. In
>> fbdev_pv_display_prepare function, I map the xen_fbfront memory to
>> qemu (fb_mem) and directly incoke ioctl(fd, IOCTL_PRIVCMD_MMAPBATCH,
>> &ioctlx) to map the vram of hvm guest.
>> 
> 
> The current implementation of IOCTL_PRIVCMD_MMAPBATCH is only capable of
> mapping foreign mfns into the address space of the caller, while we need
> to remap a set of pages already mapped in the address space of the
> caller to some gmfns of a foreign domain. In other words we need the
> same functionality but in the other direction.
> 
> First of all we need an hypercall to remap a given mfn to a guest gmfn
> and currently there are no hypercalls that do that, so we need to add a
> new one or extend an existing hypercall.
> I suggest extending xc_domain_add_to_physmap with a new XENMAPSPACE,
> called XENMAPSPACE_mfn that takes an mfn and maps it into a particular
> gmfn.

I suggest XENMAPSPACE_caller_gmfn. I'm slightly worried about the reference
counting implications of mapping these foreign pages into an HVM guest. It
might need checking with Tim Deegan.
 
> 
>> From the Xen point of view we need to add a new XENMAPSPACE_mfn case to
> xen/arch/x86/mm.c:arch_memory_op, the basic implementation should be
> something like the following (uncompiled, untested):
> 
> diff -r a79c1d5b946e xen/arch/x86/mm.c
> --- a/xen/arch/x86/mm.c Fri Aug 19 10:00:25 2011 +0100
> +++ b/xen/arch/x86/mm.c Mon Aug 22 17:51:41 2011 +0000
> @@ -4618,6 +4618,16 @@ long arch_memory_op(int op, XEN_GUEST_HA
>              page = mfn_to_page(mfn);
>              break;
>          }
> +        case XENMAPSPACE_mfn:
> +        {
> +            if ( !IS_PRIV_FOR(current->domain, d) )
> +                return -EINVAL;
> +            mfn = xatp.idx;
> +            page = mfn_to_page(mfn);

Also need something like get_page(page, current->domain).

> +            break;
> +        }
>          default:
>              break;
>          }
> @@ -4648,10 +4658,12 @@ long arch_memory_op(int op, XEN_GUEST_HA
>          }
>  
>          /* Unmap from old location, if any. */
> -        gpfn = get_gpfn_from_mfn(mfn);
> -        ASSERT( gpfn != SHARED_M2P_ENTRY );
> -        if ( gpfn != INVALID_M2P_ENTRY )
> -            guest_physmap_remove_page(d, gpfn, mfn, 0);
> +        if ( xatp.space != XENMAPSPACE_mfn && d != current->domain ) {
> +            gpfn = get_gpfn_from_mfn(mfn);
> +            ASSERT( gpfn != SHARED_M2P_ENTRY );
> +            if ( gpfn != INVALID_M2P_ENTRY )
> +                guest_physmap_remove_page(d, gpfn, mfn, 0);
> +        }
>  
>          /* Map at new location. */
>          rc = guest_physmap_add_page(d, xatp.gpfn, mfn, 0);
> 
> 
> Unfortunately qemu doesn't know how to find the mfns corresponding to
> the mmap'ed framebuffer and I would rather avoid writing a pagetable
> walker in qemu.
> Thus we need to use the linux kernel to do the virtual address to mfn
> translation and issue the hypercall.
> We can add a new privcmd ioctl IOCTL_PRIVCMD_FOREIGN_MAP: qemu calls this
> ioctl with the mmap'ed framebuffer address, the size of the framebuffer
> and the destination gmfns.
> The implementation of the ioctl in the kernel would:
> 
> - find the list of mfns corresponding to the arguments, using
> arbitrary_virt_to_machine;
> 
> - for each mfn, call the hypercall we extended with the appropriate
> arguments, it would end up being
> HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp);
> 
> - there would be no "if (!xen_initial_domain())" check, because this
> ioctl can be called by stubdoms.
> 
> 
> So the call trace would be:
> 
> qemu: ioctl(fd, IOCTL_PRIVCMD_FOREIGN_MAP, &ioctlx);
> 
> |
> v
> 
> linux: HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp);
> 
> |
> v
> 
> xen: guest_physmap_add_page
> 
> 
> Has anybody any better ideas?
> 
> 
>> diff --git a/fbdev.c b/fbdev.c
>> index a601b48..f7ff682 100644
>> --- a/fbdev.c
>> +++ b/fbdev.c
>> @@ -30,6 +30,12 @@
>>  #include "sysemu.h"
>>  #include "pflib.h"
>> 
>> +#include "hw/xen_backend.h"
>> +#include <xen/hvm/params.h>
>> +#include <sys/ioctl.h>
>> +#include "xenctrlosdep.h"
>> +#include <xen/privcmd.h>
>> +
>>  /* -------------------------------------------------------------------- */
>> 
>>  /* file handles */
>> @@ -541,29 +547,23 @@ static void fbdev_cleanup(void)
>>      }
>>  }
>> 
>> -static int fbdev_init(const char *device)
>> +static int fbdev_init(int prot, unsigned long size)
>>  {
>>      struct vt_stat vts;
>>         unsigned long page_mask;
>>      char ttyname[32];
>> 
>>      /* open framebuffer */
>> -    if (device == NULL) {
>> -       device = getenv("FRAMEBUFFER");
>> -    }
>> -    if (device == NULL) {
>> -       device = "/dev/fb0";
>> -    }
>> -    fb = open(device, O_RDWR);
>> +    fb = open("/dev/fb0", O_RDWR);
>>      if (fb == -1) {
>> -       fprintf(stderr, "open %s: %s\n", device, strerror(errno));
>> +        fprintf(stderr, "open /dev/fb0: %s\n", strerror(errno));
>>          return -1;
>>      }
>> 
>>      /* open virtual console */
>>      tty = 0;
>>      if (ioctl(tty, VT_GETSTATE, &vts) < 0) {
>>              fprintf(stderr, "Not started from virtual terminal, trying to
>> open one.\n");
>> 
>>          snprintf(ttyname, sizeof(ttyname), "/dev/tty0");
>>          tty = open(ttyname, O_RDWR);
>> @@ -632,14 +632,14 @@ static int fbdev_init(const char *device)
>>         goto err;
>>      }
>> 
>>      page_mask = getpagesize()-1;
>>      fb_mem_offset = (unsigned long)(fb_fix.smem_start) & page_mask;
>> -    fb_mem = mmap(NULL,fb_fix.smem_len+fb_mem_offset,
>> -                 PROT_READ|PROT_WRITE,MAP_SHARED,fb,0);
>> +    fb_mem = mmap(NULL, size << XC_PAGE_SHIFT, prot, MAP_SHARED, fb, 0);
>>      if (fb_mem == MAP_FAILED) {
>>         perror("mmap");
>>         goto err;
>>      }
>> +
>>      /* move viewport to upper left corner */
>>      if (fb_var.xoffset != 0 || fb_var.yoffset != 0) {
>>         fb_var.xoffset = 0;
>> @@ -930,3 +928,71 @@ void fbdev_display_uninit(void)
>>      qemu_remove_exit_notifier(&exit_notifier);
>>      uninit_mouse();
>>  }
> 
> I would rather avoid modifing fbdev_init and add a new function to
> xen-all.c that independently mmaps the xen_fbfront buffer with the right
> size and maps it into the guest.
> 
> 
>> +int fbdev_pv_display_prepare(unsigned long domid, int prot, const
>> unsigned long *arr,
>> +                                                               int *err,
>> unsigned int num)
>> +{
>> +       xen_pfn_t *pfn;
>> +       privcmd_mmapbatch_t ioctlx;
>> +       int fd,i,rc;
>> +
>> +    if (fbdev_init(prot,num) != 0) {
>> +        exit(1);
>> +    }
>> +
>> +       pfn = malloc(num * sizeof(*pfn));
>> +       if(!pfn){
>> +               errno = -ENOMEM;
>> +               rc = -1;
>> +               return rc;
>> +       }
>> +       memcpy(pfn, arr, num*sizeof(*arr));
>> +
>> +       fd = open("/proc/xen/privcmd", O_RDWR);
>> +       if(fd == -1){
>> +               fprintf(stderr,"Could not obtain handle on privcmd
>> device\n");
>> +               exit(1);
>> +       }
>> +
>> +       ioctlx.num = num;
>> +       ioctlx.dom = domid;
>> +       ioctlx.addr = (unsigned long)fb_mem;
>> +       ioctlx.arr = pfn;
>> +
>> +       rc = ioctl(fd, IOCTL_PRIVCMD_MMAPBATCH, &ioctlx);
>> +
>> +       for(i=0; i<num; i++)
>> +       {
>> +               switch(pfn[i] ^ arr[i])
>> +               {
>> +                       case 0:
>> +                               err[i] = rc != -ENOENT ? rc:0;
>> +                               continue;
>> +                       default:
>> +                               err[i] = -EINVAL;
>> +                               continue;
>> +               }
>> +               break;
>> +       }
>> +
>> +       free(pfn);
>> +
>> +       if (rc == -ENOENT && i == num)
>> +               rc=0;
>> +       else if(rc)
>> +       {
>> +               errno = -rc;
>> +               rc = -1;
>> +       }
>> +
>> +       if(rc < 0)
>> +       {
>> +               fprintf(stderr,"privcmd ioctl failed\n");
>> +               munmap(fb_mem, num << XC_PAGE_SHIFT);
>> +               return NULL;
>> +       }
>> +
>> +       close(fd);
>> +
>> +       return fb_mem;
>> +}
> 
> We shouldn't use IOCTL_PRIVCMD_MMAPBATCH, we need a new ioctl, see above.
> 
> 
>> diff --git a/hw/vga.c b/hw/vga.c
>> index 0f54734..de7dd85 100644
>> --- a/hw/vga.c
>> +++ b/hw/vga.c
>> @@ -28,6 +28,7 @@
>>  #include "vga_int.h"
>>  #include "pixel_ops.h"
>>  #include "qemu-timer.h"
>> +#include "xen_backend.h"
>> 
>>  //#define DEBUG_VGA
>>  //#define DEBUG_VGA_MEM
>> @@ -2237,7 +2238,12 @@ void vga_common_init(VGACommonState *s, int
>> vga_ram_size)
>>      s->is_vbe_vmstate = 0;
>>  #endif
>>      s->vram_offset = qemu_ram_alloc(NULL, "vga.vram", vga_ram_size);
>> +#ifdef CONFIG_STUBDOM
>> +       stubdom_vga_state = VGA_INIT_READY;
>> +#endif
>>      s->vram_ptr = qemu_get_ram_ptr(s->vram_offset);
>>      s->vram_size = vga_ram_size;
>>      s->get_bpp = vga_get_bpp;
>>      s->get_offsets = vga_get_offsets;
>> diff --git a/hw/xen_backend.c b/hw/xen_backend.c
>> index c506dfe..f4ecce4 100644
>> --- a/hw/xen_backend.c
>> +++ b/hw/xen_backend.c
>> @@ -48,6 +48,10 @@ XenGnttab xen_xcg = XC_HANDLER_INITIAL_VALUE;
>>  struct xs_handle *xenstore = NULL;
>>  const char *xen_protocol;
>> 
>> +#ifdef CONFIG_STUBDOM
>> +enum Stubdom_Vga_State stubdom_vga_state=0;
>> +#endif
>> +
>>  /* private */
>>  static QTAILQ_HEAD(XenDeviceHead, XenDevice) xendevs =
>> QTAILQ_HEAD_INITIALIZER(xendevs);
>>  static int debug = 0;
>> diff --git a/hw/xen_backend.h b/hw/xen_backend.h
>> index 3305630..36167d1 100644
>> --- a/hw/xen_backend.h
>> +++ b/hw/xen_backend.h
>> @@ -60,6 +60,16 @@ extern XenXC xen_xc;
>>  extern struct xs_handle *xenstore;
>>  extern const char *xen_protocol;
>> 
>> +#ifdef CONFIG_STUBDOM
>> +/* linux stubdom vga initialization*/
>> +enum Stubdom_Vga_State{
>> +       VGA_INIT_WAIT = 0,
>> +       VGA_INIT_READY,
>> +       VGA_INIT_COMPLETE
>> +};
>> +extern enum Stubdom_Vga_State stubdom_vga_state;
>> +#endif
>> +
>>  /* xenstore helper functions */
>>  int xenstore_write_str(const char *base, const char *node, const char *val);
>>  int xenstore_write_int(const char *base, const char *node, int ival);
>> diff --git a/xen-mapcache.c b/xen-mapcache.c
>> index 007136a..e615f98 100644
>> --- a/xen-mapcache.c
>> +++ b/xen-mapcache.c
>> @@ -20,6 +20,7 @@
>>  #include "xen-mapcache.h"
>>  #include "trace.h"
>> 
>> +#include "hw/xen.h"
>> 
>>  //#define MAPCACHE_DEBUG
>> 
>> @@ -144,8 +145,19 @@ static void xen_remap_bucket(MapCacheEntry *entry,
>>          pfns[i] = (address_index << (MCACHE_BUCKET_SHIFT-XC_PAGE_SHIFT)) +
>> i;
>>      }
>> 
>> +#ifdef CONFIG_STUBDOM
>> +       if(stubdom_vga_state == VGA_INIT_READY){
>> +               fprintf(stderr,"xen_remap_bucket: start linux stubdom
>> vga\n");
>> +               vaddr_base = fbdev_pv_display_prepare(xen_domid,
>> PROT_READ|PROT_WRITE,
>> +               
>> pfns, err, nb_pfn);
>> +               stubdom_vga_state = VGA_INIT_COMPLETE;
>> +       }else
>> +       vaddr_base = xc_map_foreign_bulk(xen_xc, xen_domid,
>> PROT_READ|PROT_WRITE,
>> +                                       pfns, err, nb_pfn);
>> +#else
>>      vaddr_base = xc_map_foreign_bulk(xen_xc, xen_domid,
>> PROT_READ|PROT_WRITE,
>>                                       pfns, err, nb_pfn);
>> +#endif
>>      if (vaddr_base == NULL) {
>>          perror("xc_map_foreign_bulk");
>>          exit(-1);
>> 
>> 
> 
> I don't like the stubdom_vga_init approach: in general it is a good idea
> to avoid state machines unless necessary.
> I would implement a new function called xen_vga_ram_map in xen-all.c
> that mmaps the xen_fbfront buffer and uses the new ioctl to
> map the buffer into the guest.
> xen_vga_ram_map should be called instead of xen_ram_alloc by
> qemu_ram_alloc_from_ptr if name == "vga.vram".
> 
> 
> Another suggestion: before starting to write any new lines of code, I
> would make sure that mmaping the /dev/fb device actually works correctly.
> For debugging purposes you can try to modify fbdev_init and write
> something to the framebuffer right after the mmap, to see if the new
> pattern is displayed correctly on the screen.
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel