Attached is the latest super pages patch with the modifications requested, the main content is consistent with Wei Huang''s previous post of this patch. applies to unstable 16877 cleanly. There is a problem that I am working that is related to c/s 16728 (see below), so the patch doesn''t function quite yet with latest unstable/staging (i.e. don''t apply to staging quite yet). Modifications from last patch: xc_hvm_build.c/setup_guest 1) all the *_2MB variables are now *_super 2) the superpage size is now calculated (currently hardcoded to PAGE_SHIFT + 9 etc) from a single variable describing the order of a superpage (21 or 22) 3) the superpage size is determined by querying the xen capabilities: if Xen supports PAE HVM guests then superpages are 2MB, else 4MB. Currently using "if ( strstr(caps, "x86_32p") )" as a string match. What is not done: 1) I don''t have crosscompile for IA64 and PowerPC compilation, I added what I think are enough modifications to allow compilation without runtime failure, but actual compiling will tell. 2) c/s 16728 Removing 16728 allows guests to start up. Get a "bad address" error currently, but removing only the xc_domain_memory_decrease_reservation(... [shared_page_nr-3]) call (which frees up the guard page), then allows HVM guests to start up. I am looking at the issue now, any ideas appreciated. Testing: 1) ran over a week with PAE x86 hv without issue on 16656 base with HVM guests. 64b hv (sans 16728) does not show any issues starting up various HVM guests with more limited testing. Used both shadow paging and nested paging during the testing. 2) in process of starting up 64b hv with patch for extended testing once issue with 16728 resolved 3) don''t plan on testing any 32b hv Jan''s recent 1Gig patch applies without failure on top of this latest patch. Please take a look at it, esp the newer mods to xc_hvm_build.c and the powerpc/ia64 areas, and provide any comments. it *should* be almost ready to apply. Cheers, Signed-off-by: Tom Woller <thomas.woller@.com> --Tom thomas.woller@amd.com +1-512-602-0059 AMD Corporation - Operating Systems Research Center 5204 E. Ben White Blvd. UBC1 Austin, Texas 78741 _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
With this patch, couldn''t create HVM guest with VT-d, prompted "Error: (14, ''Bad addrees'')". Randy (Weidong) Woller, Thomas wrote:> Attached is the latest super pages patch with the modifications > requested, the main content is consistent with Wei Huang''s previous > post of this patch. > > applies to unstable 16877 cleanly. > > There is a problem that I am working that is related to c/s 16728 (see > below), so the patch doesn''t function quite yet with latest > unstable/staging (i.e. don''t apply to staging quite yet). > > Modifications from last patch: > xc_hvm_build.c/setup_guest > 1) all the *_2MB variables are now *_super > 2) the superpage size is now calculated (currently > hardcoded to PAGE_SHIFT + 9 etc) from a single variable > describing the order of a superpage (21 or 22) > 3) the superpage size is determined by querying the xen capabilities: > if Xen supports PAE HVM guests then superpages are 2MB, else 4MB. > Currently using "if ( strstr(caps, "x86_32p") )" as a string match. > > What is not done: > 1) I don''t have crosscompile for IA64 and PowerPC compilation, I added > what I think are enough modifications to allow compilation without > runtime failure, but actual compiling will tell. > > 2) c/s 16728 > Removing 16728 allows guests to start up. Get a "bad address" error > currently, but removing only the > xc_domain_memory_decrease_reservation(... [shared_page_nr-3]) call > (which frees up the guard page), then allows HVM guests to start up. I > am looking at the issue now, any ideas appreciated. > > Testing: > 1) ran over a week with PAE x86 hv without issue on 16656 base with > HVM guests. > 64b hv (sans 16728) does not show any issues starting up various HVM > guests with more limited testing. Used both shadow paging and nested > paging during the testing. > > 2) in process of starting up 64b hv with patch for extended testing > once issue with 16728 resolved > > 3) don''t plan on testing any 32b hv > > Jan''s recent 1Gig patch applies without failure on top of this latest > patch. > > Please take a look at it, esp the newer mods to xc_hvm_build.c and the > powerpc/ia64 areas, and provide any comments. it *should* be almost > ready to apply. > > Cheers, > > Signed-off-by: Tom Woller <thomas.woller@.com> > > --Tom > > thomas.woller@amd.com +1-512-602-0059 > AMD Corporation - Operating Systems Research Center > 5204 E. Ben White Blvd. UBC1 > Austin, Texas 78741_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Woller, Thomas
2008-Jan-25 15:35 UTC
RE: [Xen-devel] [PATCH] Super Pages (2meg/4meg) patch
> > 2) c/s 16728 > > Removing 16728 allows guests to start up. Get a "bad address" error > > currently, but removing only the > > xc_domain_memory_decrease_reservation(... [shared_page_nr-3]) call > > (which frees up the guard page), then allows HVM guests to > start up. I> With this patch, couldn''t create HVM guest with VT-d, prompted "Error: > (14, ''Bad addrees'')". > > Randy (Weidong)Randy, thanks for trying this patch out. Yes, this error occurs currently with unstable, and is not a problem with VT-d specifically. This "bad address" error occurs if c/s 16728 is not removed prior to the test. Sorry about the confusion. If you remove c/s 16728 (the guard page), then the code seems to work well. I am in process of determing root cause for why the call in 16728 to: xc_domain_memory_decrease_reservation(...page_array[shared_page_nr-3]) then causes the function: xc_clear_domain_page(xc_handle, dom, shared_page_nr) To fail in the xc_map_foreign_range()/ioctl(xc_handle, IOCTL_PRIVCMD_MMAP, &ioctlx) Function w/ -EFAULT. Soon as this is fixed, I''ll repost the patch(es). tom> -----Original Message----- > From: Han, Weidong [mailto:weidong.han@intel.com] > Sent: Thursday, January 24, 2008 9:16 PM > To: Woller, Thomas; xen-devel@lists.xensource.com > Subject: RE: [Xen-devel] [PATCH] Super Pages (2meg/4meg) patch > > With this patch, couldn''t create HVM guest with VT-d, prompted "Error: > (14, ''Bad addrees'')". > > Randy (Weidong) > > Woller, Thomas wrote: > > Attached is the latest super pages patch with the modifications > > requested, the main content is consistent with Wei Huang''s previous > > post of this patch. > > > > applies to unstable 16877 cleanly. > > > > There is a problem that I am working that is related to c/s > 16728 (see > > below), so the patch doesn''t function quite yet with latest > > unstable/staging (i.e. don''t apply to staging quite yet). > > > > Modifications from last patch: > > xc_hvm_build.c/setup_guest > > 1) all the *_2MB variables are now *_super > > 2) the superpage size is now calculated (currently > > hardcoded to PAGE_SHIFT + 9 etc) from a single variable > > describing the order of a superpage (21 or 22) > > 3) the superpage size is determined by querying the xen > capabilities: > > if Xen supports PAE HVM guests then superpages are 2MB, else 4MB. > > Currently using "if ( strstr(caps, "x86_32p") )" as a > string match. > > > > What is not done: > > 1) I don''t have crosscompile for IA64 and PowerPC > compilation, I added > > what I think are enough modifications to allow compilation without > > runtime failure, but actual compiling will tell. > > > > 2) c/s 16728 > > Removing 16728 allows guests to start up. Get a "bad address" error > > currently, but removing only the > > xc_domain_memory_decrease_reservation(... [shared_page_nr-3]) call > > (which frees up the guard page), then allows HVM guests to > start up. I > > am looking at the issue now, any ideas appreciated. > > > > Testing: > > 1) ran over a week with PAE x86 hv without issue on 16656 base with > > HVM guests. > > 64b hv (sans 16728) does not show any issues starting up > various HVM > > guests with more limited testing. Used both shadow paging > and nested > > paging during the testing. > > > > 2) in process of starting up 64b hv with patch for extended testing > > once issue with 16728 resolved > > > > 3) don''t plan on testing any 32b hv > > > > Jan''s recent 1Gig patch applies without failure on top of > this latest > > patch. > > > > Please take a look at it, esp the newer mods to > xc_hvm_build.c and the > > powerpc/ia64 areas, and provide any comments. it *should* > be almost > > ready to apply. > > > > Cheers, > > > > Signed-off-by: Tom Woller <thomas.woller@.com> > > > > --Tom > > > > thomas.woller@amd.com +1-512-602-0059 AMD Corporation - Operating > > Systems Research Center > > 5204 E. Ben White Blvd. UBC1 > > Austin, Texas 78741 > > > >_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel