Re: [Xen-devel] xl: pci completion error

[Sergey Tovpeko <tsv.devel@xxxxxxxxx>]

Hello, Gianni!

I was interested, how your patch would behave with many pci devices 
given to a domain.
It seemed that each device removal takes 10 seconds  for the timeout.

pci=[ '01:00.0','00:1d.0', '00:1d.1', '00:1d.2','00:1d.3', '00:1d.7'  ]


The result was a bit different. After the first pci removal,  the others 
somehow were destroyed.
So the second do_pci_remove run into the failure that there were no pci 
devices at all.
So I got the following log.

Waiting for domain workxp (domid 1) to die [pid 3900]
Domain 1 is dead
Action for shutdown reason code 0 is destroy
Domain 1 needs to be cleaned up: destroying the domain
do_pci_remove device 01:00.0
libxl: error: libxl_device.c:448:libxl__wait_for_device_model Device 
Model not ready
libxl: error: libxl_pci.c:861:do_pci_remove Device Model didn't respond 
in time
do_pci_remove device 00:1d.0
libxl: error: libxl_pci.c:839:do_pci_remove PCI device not attached to 
this domain
libxl: error: libxl.c:944:libxl_domain_destroy pci shutdown failed for 
domid 1
libxl: error: libxl.c:896:libxl_destroy_device_model Couldn't find 
device model's pid: No such file or directory
libxl: error: libxl.c:956:libxl_domain_destroy 
libxl_destroy_device_model failed for 1
libxl: error: libxl_device.c:307:libxl__devices_destroy 
/local/domain/1/device is empty

Sergey.

> Sergeys analysis sounds very plausible to me actually. The co-ordination
> required between qemu and libxl for PCI passthrough is very complicated
> and one ought to have fairly low confidence in it's correctness :P
>
> Below is a less hacky version of the patch I just sent. Stefano, please
> consider this for inclusion.
>
> ---
> xl: Implement PCI passthrough force removal
>
> This fixes two errors with removing PCI devices from HVM domains. The
> first error is that the handling of "pci-rem" device-model command is
> erroneously implemented in qemu and difficult (impossible?) to get
> right.
>
> For example, during domain shutdown there can be a race where the guest
> OS unloads it's drivers and perhaps even shuts down PCI subsystem before
> the pci-rem command has been received by qemu. This means that no OS is
> present to write to the port which causes the dm command to be
> acknowledged.
>
> We fix this by implementing a 'force removal' option to
> libxl_device_pci_remove which is always set to 1 during guest shutdown.
> It can be optionally enabled on the xl command line for other occasions.
>
> The second error is that if a guest OS doesn't respond to the SCI
> interrupt and therefore the pci-rem dm command, which can happen if the
> guest OS has no ACPI PCI hotplug support, then device removal bails with
> an error but only AFTER removing the device from xenstore. This means
> that xenstore gets in to an inconsistent state where an assigned device
> also appears to be assignable.
>
> This is fixed by moving xenstore device removal to occur only after the
> device has really been removed.
>
> Signed-off-by: Gianni Tedesco <gianni.tedesco@xxxxxxxxxx>
>
> diff -r 02e199c96ece tools/libxl/libxl.h
> --- a/tools/libxl/libxl.h       Wed Oct 06 11:00:19 2010 +0100
> +++ b/tools/libxl/libxl.h       Wed Oct 06 15:29:12 2010 +0100
> @@ -406,7 +406,7 @@ int libxl_device_vfb_clean_shutdown(libx
>  int libxl_device_vfb_hard_shutdown(libxl_ctx *ctx, uint32_t domid);
>  
>  int libxl_device_pci_add(libxl_ctx *ctx, uint32_t domid, libxl_device_pci *pcidev);
> -int libxl_device_pci_remove(libxl_ctx *ctx, uint32_t domid, libxl_device_pci 
> *pcidev);
> +int libxl_device_pci_remove(libxl_ctx *ctx, uint32_t domid, libxl_device_pci 
> *pcidev, int force);
>  int libxl_device_pci_shutdown(libxl_ctx *ctx, uint32_t domid);
>  int libxl_device_pci_list_assigned(libxl_ctx *ctx, libxl_device_pci **list, 
> uint32_t domid, int *num);
>  int libxl_device_pci_list_assignable(libxl_ctx *ctx, libxl_device_pci **list, 
> int *num);
> diff -r 02e199c96ece tools/libxl/libxl_pci.c
> --- a/tools/libxl/libxl_pci.c   Wed Oct 06 11:00:19 2010 +0100
> +++ b/tools/libxl/libxl_pci.c   Wed Oct 06 15:29:12 2010 +0100
> @@ -841,7 +841,8 @@ out:
>      return rc;
>  }
>  
> -static int do_pci_remove(libxl__gc *gc, uint32_t domid, libxl_device_pci *pcidev)
> +static int do_pci_remove(libxl__gc *gc, uint32_t domid,
> +                         libxl_device_pci *pcidev, int force)
>  {
>      libxl_ctx *ctx = libxl__gc_owner(gc);
>      libxl_device_pci *assigned;
> @@ -858,8 +859,6 @@ static int do_pci_remove(libxl__gc *gc, 
>          }
>      }
>  
> -    libxl_device_pci_remove_xenstore(gc, domid, pcidev);
> -
>      hvm = libxl__domain_is_hvm(ctx, domid);
>      if (hvm) {
>          if (libxl__wait_for_device_model(ctx, domid, "running", NULL, NULL) < 
> 0) {
> @@ -878,7 +877,12 @@ static int do_pci_remove(libxl__gc *gc, 
>              xs_write(ctx->xsh, XBT_NULL, path, "pci-rem", strlen("pci-rem"));
>              if (libxl__wait_for_device_model(ctx, domid, "pci-removed", NULL, 
> NULL) < 0) {
>                  LIBXL__LOG(ctx, LIBXL__LOG_ERROR, "Device Model didn't respond in 
> time");
> -                return ERROR_FAIL;
> +                /* This depends on guest operating system acknowledging the
> +                 * SCI, if it doesn't respond in time then we may wish to 
> +                 * force the removal.
> +                 */
> +                if ( !force )
> +                    return ERROR_FAIL;
>              }
>          }
>          path = libxl__sprintf(gc, "/local/domain/0/device-model/%d/state", 
> domid);
> @@ -924,7 +928,7 @@ skip1:
>          if ((fscanf(f, "%u", &irq) == 1) && irq) {
>              rc = xc_physdev_unmap_pirq(ctx->xch, domid, irq);
>              if (rc < 0) {
> -                LIBXL__LOG_ERRNOVAL(ctx, LIBXL__LOG_ERROR, rc, "xc_physdev_map_pirq 
> irq=%d", irq);
> +                LIBXL__LOG_ERRNOVAL(ctx, LIBXL__LOG_ERROR, rc, 
> "xc_physdev_unmap_pirq irq=%d", irq);
>              }
>              rc = xc_domain_irq_permission(ctx->xch, domid, irq, 0);
>              if (rc < 0) {
> @@ -948,13 +952,16 @@ out:
>      stubdomid = libxl_get_stubdom_id(ctx, domid);
>      if (stubdomid != 0) {
>          libxl_device_pci pcidev_s = *pcidev;
> -        libxl_device_pci_remove(ctx, stubdomid, &pcidev_s);
> +        libxl_device_pci_remove(ctx, stubdomid, &pcidev_s, force);
>      }
>  
> +    libxl_device_pci_remove_xenstore(gc, domid, pcidev);
> +
>      return 0;
>  }
>  
> -int libxl_device_pci_remove(libxl_ctx *ctx, uint32_t domid, libxl_device_pci *pcidev)
> +int libxl_device_pci_remove(libxl_ctx *ctx, uint32_t domid,
> +                            libxl_device_pci *pcidev, int force)
>  {
>      libxl__gc gc = LIBXL_INIT_GC(ctx);
>      unsigned int orig_vdev, pfunc_mask;
> @@ -980,7 +987,7 @@ int libxl_device_pci_remove(libxl_ctx *c
>              }else{
>                  pcidev->vdevfn = orig_vdev;
>              }
> -            if ( do_pci_remove(&gc, domid, pcidev) )
> +            if ( do_pci_remove(&gc, domid, pcidev, force) )
>                  rc = ERROR_FAIL;
>          }
>      }
> @@ -1049,7 +1056,11 @@ int libxl_device_pci_shutdown(libxl_ctx 
>      if ( rc )
>          return rc;
>      for (i = 0; i < num; i++) {
> -        if (libxl_device_pci_remove(ctx, domid, pcidevs + i) < 0)
> +        /* Force remove on shutdown since, on HVM, qemu will not always
> +         * respond to SCI interrupt because the guest kernel has shut down the
> +         * devices by the time we even get here!
> +         */
> +        if (libxl_device_pci_remove(ctx, domid, pcidevs + i, 1) < 0)
>              return ERROR_FAIL;
>      }
>      free(pcidevs);
> diff -r 02e199c96ece tools/libxl/xl_cmdimpl.c
> --- a/tools/libxl/xl_cmdimpl.c  Wed Oct 06 11:00:19 2010 +0100
> +++ b/tools/libxl/xl_cmdimpl.c  Wed Oct 06 15:29:12 2010 +0100
> @@ -2262,7 +2262,7 @@ int main_pcilist(int argc, char **argv)
>      return 0;
>  }
>  
> -static void pcidetach(char *dom, char *bdf)
> +static void pcidetach(char *dom, char *bdf, int force)
>  {
>      libxl_device_pci pcidev;
>  
> @@ -2273,20 +2273,24 @@ static void pcidetach(char *dom, char *b
>          fprintf(stderr, "pci-detach: malformed BDF specification \"%s\"\n", 
> bdf);
>          exit(2);
>      }
> -    libxl_device_pci_remove(&ctx, domid, &pcidev);
> +    libxl_device_pci_remove(&ctx, domid, &pcidev, force);
>      libxl_device_pci_destroy(&pcidev);
>  }
>  
>  int main_pcidetach(int argc, char **argv)
>  {
>      int opt;
> +    int force = 0;
>      char *domname = NULL, *bdf = NULL;
>  
> -    while ((opt = getopt(argc, argv, "h")) != -1) {
> +    while ((opt = getopt(argc, argv, "hf")) != -1) {
>          switch (opt) {
>          case 'h':
>              help("pci-detach");
>              return 0;
> +        case 'f':
> +            force = 1;
> +            break;
>          default:
>              fprintf(stderr, "option not supported\n");
>              break;
> @@ -2300,7 +2304,7 @@ int main_pcidetach(int argc, char **argv
>      domname = argv[optind];
>      bdf = argv[optind + 1];
>  
> -    pcidetach(domname, bdf);
> +    pcidetach(domname, bdf, force);
>      return 0;
>  }
>  static void pciattach(char *dom, char *bdf, char *vs)
>
>
>   

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel