2:54 p.m.
On Thu, Jan 09, 2014 at 01:21:35PM -0500, Dave Anderson wrote:
----- Original Message -----
> On Thu, Jan 09, 2014 at 11:40:59AM -0500, Dave Anderson wrote:
> >
> >
> > ----- Original Message -----
> > >
> > > Version 2
> > > - Moves the warning to this point:
> > > ...
> > > This GDB was configured as "x86_64-unknown-linux-gnu"...
> > >
> > > WARNING: All unused vmemmap page structures are excluded from this
> > > dump.
> > > This will cause failures of the kmem command if it
> > > attempts to
> > > walk any list of free pages (with options -f -F -i -s and
> > > -S).
> > >
> > > SYSTEM MAP: /boot/System.map-2.6.32-cpw
> > > ...
> > >
> > > - Drop patch 2, which added warnings to individual kmem options.
> > >
> > > - Feel free to change the wording of the warning.
> > >
> > > And this patch is contingent upon the acceptance of a change to the
> > > makedumpfile command.
> > > http://marc.info/?l=kexec&m=138853299130125&w=2
> > >
> > >
> > > If makedumpfile excludes unused page structures it will flag that
> > > fact in the dump header.
> > > (There are about 3.67 million pages full of page structures for
> > > every tera byte of system memory. The great bulk of those
> > > page structures are not needed.)
> > > Their exclusion is a makedumpfile option.
> > >
> > > Crash will display a note during initialization if such structures
> > > have been excluded. Crash commands that walk page freelists, for
> > > example, will fail. So the note will help the user understand why.
> > >
> > > Signed-off-by: Cliff Wickman <cpw(a)sgi.com>
> > > ---
> > > defs.h | 1 +
> > > diskdump.c | 3 +++
> > > diskdump.h | 1 +
> > > main.c | 6 ++++++
> > > 4 files changed, 11 insertions(+)
> > >
> > > Index: crash-7.0.4/diskdump.c
> > > ===================================================================
> > > --- crash-7.0.4.orig/diskdump.c
> > > +++ crash-7.0.4/diskdump.c
> > > @@ -749,6 +749,9 @@ restart:
> > > dd->valid_pages[i]++;
> > > }
> > >
> > > + if (header->status & DUMP_DH_EXCLUDED_VMEMMAP)
> > > + pc->flags2 |= VMEXCLUDED;
> > > +
> > > return TRUE;
> > >
> > > err:
> > > Index: crash-7.0.4/diskdump.h
> > > ===================================================================
> > > --- crash-7.0.4.orig/diskdump.h
> > > +++ crash-7.0.4/diskdump.h
> > > @@ -84,6 +84,7 @@ struct kdump_sub_header {
> > > #define DUMP_DH_COMPRESSED_ZLIB 0x1 /* page is compressed with zlib
> > > */
> > > #define DUMP_DH_COMPRESSED_LZO 0x2 /* page is compressed with lzo
> > > */
> > > #define DUMP_DH_COMPRESSED_SNAPPY 0x4 /* page is compressed with
> > > snappy
> > > */
> > > +#define DUMP_DH_EXCLUDED_VMEMMAP 0x8 /* unused vmemmap pages are
> > > excluded */
> > >
> > > /* descriptor of each page for vmcore */
> > > typedef struct page_desc {
> > > Index: crash-7.0.4/defs.h
> > > ===================================================================
> > > --- crash-7.0.4.orig/defs.h
> > > +++ crash-7.0.4/defs.h
> > > @@ -505,6 +505,7 @@ struct program_context {
> > > #define VMCOREINFO (0x400ULL)
> > > #define ALLOW_FP (0x800ULL)
> > > #define REM_PAUSED_F (0x1000ULL)
> > > +#define VMEXCLUDED (0x2000ULL)
> > > #define REMOTE_PAUSED() (pc->flags2 & REM_PAUSED_F)
> > > char *cleanup;
> > > char *namelist_orig;
> > > Index: crash-7.0.4/main.c
> > > ===================================================================
> > > --- crash-7.0.4.orig/main.c
> > > +++ crash-7.0.4/main.c
> > > @@ -662,6 +662,12 @@ main_loop(void)
> > > } else
> > > SIGACTION(SIGINT, restart, &pc->sigaction, NULL);
> > >
> > > + if (pc->flags2 & VMEXCLUDED)
> > > + fprintf(fp,
> > > + "WARNING: All unused vmemmap page structures are excluded
from
> > > this dump.\n"
> > > + " This will cause failures of the kmem command if it
attempts
> > > to\n"
> > > + " walk any list of free pages (with options -f -F -i -s
and
> > > -S).\n\n");
> > > +
> > > /*
> > > * Display system statistics and current context.
> > > */
> > >
> >
> > This patch looks reasonable.
> >
> > BTW, what happens if you enter "kmem <address>" alone with no
option?
> > It should fail as well, no?
>
> Yes, it does.
>
> crash> kmem 133360
> kmem: page excluded: kernel virtual address: ffffea0000004350 type:
> "page.lru.next"
>
> crash> kmem ffff8897baae7540
> CACHE NAME OBJSIZE ALLOCATED TOTAL SLABS
> SSIZE
> ffff88c7bec70040 task_struct 2656 4578 5361 1787
> 8k
> SLAB MEMORY TOTAL ALLOCATED FREE
> ffff8897baae6040 ffff8897baae6080 3 2 1
> FREE / [ALLOCATED]
> [ffff8897baae7540]
>
> kmem: page excluded: kernel virtual address: ffffea0000007028 type: "first
> list entry"
>
> Another way to search a freelist, that I didn't notice.
> Do you want to add that to the warning?
I'm thinking that there might be others as well. Like how does "kmem -p"
handle
the missing page structs?
kmem -p gives the appearance of working, but is unable to read the
mem_map pages. It is giving false results.
This is an example of what you feared.
If we do this:
---
memory.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
Index: crash-7.0.4/memory.c
===================================================================
--- crash-7.0.4.orig/memory.c
+++ crash-7.0.4/memory.c
@@ -5657,7 +5657,7 @@ fill_mem_map_cache(ulong pp, ulong ppend
* Try to read it in one fell swoop.
*/
if (readmem(pp, KVADDR, page_cache, SIZE(page) * PGMM_CACHED,
- "page struct cache", RETURN_ON_ERROR|QUIET))
+ "page struct cache", FAULT_ON_ERROR))
return;
/*
Then we see the immediate failure:
crash> kmem -p
PAGE PHYSICAL MAPPING INDEX CNT FLAGS
kmem: page excluded: kernel virtual address: ffffea0000004000 type: "page struct
cache"
What should be done about that?
I wonder why the reads of the mem_map were done with RETURN_ON_ERROR|QUIET
to begin with?
-Cliff
It might be better to just generally warn that results may be unpredictable
because any command that accesses page structures, such as kmem, may fail.
By listing specific options, it makes it sound like that they are the only
ones, so it may be better to leave it open-ended.
Dave
--
Cliff Wickman
SGI
cpw(a)sgi.com
(651) 683-3824