File libxl.introduce-an-option-to-disable-the-non-O_... of Package xen

Overview Repositories Revisions Requests Users Attributes Meta

File libxl.introduce-an-option-to-disable-the-non-O_DIRECT-workaround.patch of Package xen (Revision 078ee701765e3bda05366c837b94d8b7)

Currently displaying revision 078ee701765e3bda05366c837b94d8b7 , Show latest

xen-unstable commit 6ec48cf41b6656c98148380f39010063e62628c5
Frp,: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Date:   Wed Apr 30 16:06:24 2014 +0100
Subject: libxl: introduce an option for disabling the non-O_DIRECT workaround
    
Document and implement a new option that permits disk backends which
would otherwise have to avoid O_DIRECT (because of the network memory
lifetime bug) to use it anyway.  This is:
 direct-io-safe   in the xl domain disk config specification
 direct_io_safe   in the libxl disk API
 direct-io-safe   in the backend xenstore interface

Add a reference to xen/include/public/io/blkif.h in
docs/misc/vbd-interface.txt.

This change does not break ABI. Instead of adding a new member
direct_io_safe to struct libxl_device_disk the existing readwrite member
is reused.

Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Tested-by: Felipe Franciosi <felipe@paradoxo.org>
---
 docs/misc/vbd-interface.txt         |    6 +++++
 docs/misc/xl-disk-configuration.txt |   38 ++++++++++++++++++++++++++++++++++++
 tools/libxl/libxl.c                 |    2 +
 tools/libxl/libxl.h                 |   11 ++++++++++
 tools/libxl/libxlu_disk.c           |    2 +
 tools/libxl/libxlu_disk_i.h         |    2 -
 tools/libxl/libxlu_disk_l.l         |    1 
 xen/include/public/io/blkif.h       |   22 ++++++++++++++++++++
 8 files changed, 83 insertions(+), 1 deletion(-)

Index: xen-4.4.4-testing/docs/misc/vbd-interface.txt
===================================================================
--- xen-4.4.4-testing.orig/docs/misc/vbd-interface.txt
+++ xen-4.4.4-testing/docs/misc/vbd-interface.txt
@@ -125,3 +125,9 @@ because they directly map the bottom 8 b
 directly to the Linux guest's device number and throw away the rest;
 they can crash due to minor number clashes.  With these guests, the
 workaround is not to supply problematic combinations of devices.
+
+
+Other frontend and backend options
+----------------------------------
+
+See xen/include/public/io/blkif.h for the full list of options.
Index: xen-4.4.4-testing/docs/misc/xl-disk-configuration.txt
===================================================================
--- xen-4.4.4-testing.orig/docs/misc/xl-disk-configuration.txt
+++ xen-4.4.4-testing/docs/misc/xl-disk-configuration.txt
@@ -178,6 +178,44 @@ information to be interpreted by the exe
 These scripts are normally called "block-<script>".
 
 
+direct-io-safe
+--------------
+
+Description:           Disables non-O_DIRECT workaround
+Supported values:      absent, present
+Mandatory:             No
+Default value:         absent (workaround may be enabled)
+
+There is a memory lifetime bug in some driver domain (dom0) kernels
+which can cause crashes when using O_DIRECT.  The bug occurs due to a
+mismatch between the backend-visible lifetime of pages used for the
+Xen PV network protocol and that expected by the backend kernel's
+networking subsystem.  This can cause crashes when using certain
+backends with certain underlying storage.
+
+See:
+ http://lists.xen.org/archives/html/xen-devel/2012-12/msg01154.html
+
+For this reason, (this version of) the Xen libxl toolstack disables
+O_DIRECT when using the qemu-based Xen PV backend ("qdisk").
+
+However, this workaround has performance and scaling implications, and
+it is only necessary if the underlying device is a network filesystem.
+If the underlying device is not, then it is good to disable it; that
+is what this option is for.
+
+This option simply requests that the workaround be disabled.  (However,
+not all backends versions which use the workaround understand this
+option, so this is on a best effort basis.)
+
+It's important to note that if you are storing the VM disk on a
+network filesystem or a network block device (NFS or ISCSI) it might
+not be safe to use this option.  Otherwise specifying it is safe and
+can give better performances.
+
+If in the future the bug is fixed properly this option will then be
+silently ignored.
+
 
 discard / no-discard
 ---------------
Index: xen-4.4.4-testing/tools/libxl/libxl.c
===================================================================
--- xen-4.4.4-testing.orig/tools/libxl/libxl.c
+++ xen-4.4.4-testing/tools/libxl/libxl.c
@@ -2519,6 +2519,8 @@ static void device_disk_add(libxl__egc *
         flexarray_append(back, disk->readwrite ? "w" : "r");
         flexarray_append(back, "device-type");
         flexarray_append(back, disk->is_cdrom ? "cdrom" : "disk");
+        if ((disk->readwrite & ~LIBXL_HAVE_LIBXL_DEVICE_DISK_DIRECT_IO_SAFE_MASK) == LIBXL_HAVE_LIBXL_DEVICE_DISK_DIRECT_IO_SAFE_MAGIC)
+            flexarray_append_pair(back, "direct-io-safe", "1");
         if ((disk->readwrite & ~LIBXL_HAVE_LIBXL_DEVICE_DISK_DISCARD_DISABLE_MASK) == LIBXL_HAVE_LIBXL_DEVICE_DISK_DISCARD_DISABLE_MAGIC)
             flexarray_append_pair(back, "discard-enable", "0");
 
Index: xen-4.4.4-testing/tools/libxl/libxl.h
===================================================================
--- xen-4.4.4-testing.orig/tools/libxl/libxl.h
+++ xen-4.4.4-testing/tools/libxl/libxl.h
@@ -95,6 +95,17 @@
 #define LIBXL_HAVE_BUILDINFO_EVENT_CHANNELS 1
 
 /*
+ * The libxl_device_disk lacks some "cache" field, enabling directio
+ * is supported without breaking the ABI. This is done by overloading
+ * struct libxl_device_disk->readwrite:
+ * readwrite == 0: disk is readonly, no directio
+ * readwrite == 1: disk is readwrite, backend driver may enable directio
+ * readwrite == MAGIC: disk is readwrite, backend driver should use direct IO
+ */
+#define LIBXL_HAVE_LIBXL_DEVICE_DISK_DIRECT_IO_SAFE_MAGIC 0x00000600U
+#define LIBXL_HAVE_LIBXL_DEVICE_DISK_DIRECT_IO_SAFE_MASK  0xfffff0ffU
+
+/*
  * The libxl_device_disk lacks discard_enable field, disabling discard
  * is supported without breaking the ABI. This is done by overloading
  * struct libxl_device_disk->readwrite:
Index: xen-4.4.4-testing/tools/libxl/libxlu_disk.c
===================================================================
--- xen-4.4.4-testing.orig/tools/libxl/libxlu_disk.c
+++ xen-4.4.4-testing/tools/libxl/libxlu_disk.c
@@ -80,6 +80,8 @@ int xlu_disk_parse(XLU_Config *cfg,
             disk->format = LIBXL_DISK_FORMAT_EMPTY;
     }
 
+    if (disk->readwrite && dpc.direct_io_safe)
+        disk->readwrite = (disk->readwrite & LIBXL_HAVE_LIBXL_DEVICE_DISK_DIRECT_IO_SAFE_MASK) | LIBXL_HAVE_LIBXL_DEVICE_DISK_DIRECT_IO_SAFE_MAGIC;
     if (disk->readwrite && dpc.disable_discard)
         disk->readwrite = (disk->readwrite & LIBXL_HAVE_LIBXL_DEVICE_DISK_DISCARD_DISABLE_MASK) | LIBXL_HAVE_LIBXL_DEVICE_DISK_DISCARD_DISABLE_MAGIC;
 
Index: xen-4.4.4-testing/tools/libxl/libxlu_disk_i.h
===================================================================
--- xen-4.4.4-testing.orig/tools/libxl/libxlu_disk_i.h
+++ xen-4.4.4-testing/tools/libxl/libxlu_disk_i.h
@@ -10,7 +10,7 @@ typedef struct {
     void *scanner;
     YY_BUFFER_STATE buf;
     libxl_device_disk *disk;
-    int access_set, disable_discard, had_depr_prefix;
+    int access_set, direct_io_safe, disable_discard, had_depr_prefix;
     const char *spec;
 } DiskParseContext;
 
Index: xen-4.4.4-testing/tools/libxl/libxlu_disk_l.l
===================================================================
--- xen-4.4.4-testing.orig/tools/libxl/libxlu_disk_l.l
+++ xen-4.4.4-testing/tools/libxl/libxlu_disk_l.l
@@ -173,6 +173,7 @@ backendtype=[^,]*,? { STRIP(','); setbac
 
 vdev=[^,]*,?	{ STRIP(','); SAVESTRING("vdev", vdev, FROMEQUALS); }
 script=[^,]*,?	{ STRIP(','); SAVESTRING("script", script, FROMEQUALS); }
+direct-io-safe,? { DPC->direct_io_safe = 1; }
 discard,?	{ DPC->disable_discard = 0; }
 no-discard,?	{ DPC->disable_discard = 1; }
 
Index: xen-4.4.4-testing/xen/include/public/io/blkif.h
===================================================================
--- xen-4.4.4-testing.orig/xen/include/public/io/blkif.h
+++ xen-4.4.4-testing/xen/include/public/io/blkif.h
@@ -97,6 +97,28 @@
  *
  *      The type of the backing device/object.
  *
+ *
+ * direct-io-safe
+ *      Values:         0/1 (boolean)
+ *      Default Value:  0
+ *
+ *      The underlying storage is not affected by the direct IO memory
+ *      lifetime bug.  See:
+ *        http://lists.xen.org/archives/html/xen-devel/2012-12/msg01154.html
+ *
+ *      Therefore this option gives the backend permission to use
+ *      O_DIRECT, notwithstanding that bug.
+ *
+ *      That is, if this option is enabled, use of O_DIRECT is safe,
+ *      in circumstances where we would normally have avoided it as a
+ *      workaround for that bug.  This option is not relevant for all
+ *      backends, and even not necessarily supported for those for
+ *      which it is relevant.  A backend which knows that it is not
+ *      affected by the bug can ignore this option.
+ *
+ *      This option doesn't require a backend to use O_DIRECT, so it
+ *      should not be used to try to control the caching behaviour.
+ *
  *--------------------------------- Features ---------------------------------
  *
  * feature-barrier

Places

File libxl.introduce-an-option-to-disable-the-non-O_DIRECT-workaround.patch of Package xen (Revision 078ee701765e3bda05366c837b94d8b7)

Places