lsmem/chmem: add memory zone awareness

These patches are against lsmem/chmem in util-linux, they add support for listing and changing memory zone allocation.

Added Michal Hocko and linux-mm on cc, to raise general awareness for the lsmem/chmem tools, and the new memory zone functionality in particular. I think this can be quite useful for memory hotplug kernel development, and if not, sorry for the noise.

Andre Wild (1):
lsmem/chmem: add memory zone awareness to bash-completion

Gerald Schaefer (2):
lsmem/chmem: add memory zone awareness tests/lsmem: update lsmem test with ZONES column

 bash-completion/chmem                  |   1 +
 bash-completion/lsmem                  |   2 +-
 sys-utils/chmem.8                      |  19 +++++
 sys-utils/chmem.c                      | 136 +++++++++++++++++++++++++++++++--
 sys-utils/lsmem.1                      |   4 +-
 sys-utils/lsmem.c                      |  98 +++++++++++++++++++++++-
 tests/expected/lsmem/lsmem-s390-zvm-6g |  21 +++++
 tests/expected/lsmem/lsmem-x86_64-16g  |  39 ++++++++++
 tests/ts/lsmem/lsmem                   |   1 +
 9 files changed, 309 insertions(+), 12 deletions(-)

2.13.5

With this patch, valid memory zones can be shown with lsmem, and chmem can set memory online/offline in a specific memory zone, if allowed by the kernel. The valid memory zones are read from the "valid_zones" sysfs attribute, and setting memory online to a specific zone is done by echoing "online_kernel" or "online_movable" to the "state" sysfs attribute, in addition to the previous "online".

This patch also changes the default behavior of chmem, when setting memory online without specifying a memory zone. If valid, memory will be set online to the zone Movable. This zone is preferable for memory hotplug, as it makes memory offline much more likely to succeed.

---
 sys-utils/chmem.8 |  19 ++++++++
 sys-utils/chmem.c | 136 ++++++++++++++++++++++++++++++++++++++++++++++++++----
 sys-utils/lsmem.1 |   4 +-
 sys-utils/lsmem.c |  98 ++++++++++++++++++++++++++++++++++++++-
 4 files changed, 246 insertions(+), 11 deletions(-)

index a116bc9e7a4b..dae7413d4899 100644
@@ -5,6 +5,7 @@ chmem \- configure memory
 .B chmem
 .RB [ \-h "] [" \-V "] [" \-v "] [" \-e | \-d "]"
 [\fISIZE\fP|\fIRANGE\fP|\fB\-b\fP \fIBLOCKRANGE\fP]
+[-z ZONE]
 .SH DESCRIPTION
 The chmem command sets a particular size or range of memory online or offline.
 .

@@ -25,6 +26,19 @@ and <last> is the number of the last memory block in the memory
 range. Alternatively a single block can be specified. \fIBLOCKRANGE\fP requires
 the \fB--blocks\fP option.
 .
+.IP "\(hy" 2
+Specify \fIZONE\fP as the name of a memory zone, as shown in the output of the
+\fBlsmem -o +ZONES\fP command. The output shows one or more valid memory zones
+for each memory range. If multiple zones are shown, then the memory range
+currently belongs to the first zone. By default, chmem will set memory online
+to the zone Movable, if this is among the valid zones. This default can be
+changed by specifying the \fB--zone\fP option with another valid zone.
+For memory ballooning, it is recommended to select the zone Movable for memory
+online and offline, if possible. Memory in this zone is much more likely to be
+able to be offlined again, but it cannot be used for arbitrary kernel
+allocations, only for migratable pages (e.g. anonymous and page cache pages).
+Use the \fB\-\-help\fR option to see all available zones.
+.
 .PP
 \fISIZE\fP and \fIRANGE\fP must be aligned to the Linux memory block size, as
 shown in the output of the \fBlsmem\fP command.

@@ -51,6 +65,11 @@ Set the specified \fIRANGE\fP, \fISIZE\fP, or \fIBLOCKRANGE\fP of memory offline
 .BR \-e ", " \-\-enable
 Set the specified \fIRANGE\fP, \fISIZE\fP, or \fIBLOCKRANGE\fP of memory online.
 .TP
+.BR \-z ", " \-\-zone
+Select the memory \fIZONE\fP where to set the specified \fIRANGE\fP, \fISIZE\fP,
+or \fIBLOCKRANGE\fP of memory online or offline. By default, memory will be set
+online to the zone Movable, if possible.
+.TP
 .BR \-h ", " \-\-help
 Print a short help text, then exit.
 .TP

index be2862d9476a..e7df50a4ebff 100644
@@ -41,12 +41,12 @@ Do not print a header line.
 .BR \-o , " \-\-output " \fIlist\fP
 Specify which output columns to print.  Use \fB\-\-help\fR
 to get a list of all supported columns.
+The default list of columns may be extended if \fIlist\fP is
+specified in the format \fB+\fIlist\fP (e.g. \fBlsmem \-o +NODE\fP).
 .TP
 .BR \-P , " \-\-pairs"
 Produce output in the form of key="value" pairs.
 All potentially unsafe characters are hex-escaped (\\x<code>).
-The default list of columns may be extended if \fIlist\fP is
-specified in the format \fB+\fIlist\fP (e.g. \fBlsmem \-o +NODE\fP).
 .TP
 .BR \-r , " \-\-raw"
 Produce output in raw format.  All potentially unsafe characters are hex-escaped

2.13.5

The existing s390 and x86_64 dumps already contain the valid_zones sysfs attribute, so just add a new "lsmem -o +ZONES" test command and update the expected results.

---
 tests/expected/lsmem/lsmem-s390-zvm-6g | 21 ++++++++++++++++++
 tests/expected/lsmem/lsmem-x86_64-16g  | 39 ++++++++++++++++++++++++++++++++++
 tests/ts/lsmem/lsmem                   |  1 +
 3 files changed, 61 insertions(+)

index 05af40d4df9f..9f4c805ad0a6 100644
@@ -106,3 +106,24 @@ $ lsmem --json --output RANGE,SIZE,STATE,REMOVABLE,BLOCK,NODE
       {"range": "0x0000000140000000-0x000000017fffffff", "size": "1G", "state": "offline", "removable": "-", "block": "20-23", "node": "0"}
    ]
 }
+
+---
+
+$ lsmem -o +ZONES
+RANGE                                  SIZE   STATE REMOVABLE BLOCK          ZONES
+0x0000000000000000-0x000000006fffffff  1.8G  online       yes   0-6            DMA
+0x0000000070000000-0x000000007fffffff  256M  online        no     7     DMA/Normal
+0x0000000080000000-0x000000009fffffff  512M  online       yes   8-9         Normal
+0x00000000a0000000-0x00000000bfffffff  512M  online        no 10-11         Normal
+0x00000000c0000000-0x00000000dfffffff  512M  online       yes 12-13         Normal
+0x00000000e0000000-0x00000000efffffff  256M offline         -    14         Normal
+0x00000000f0000000-0x00000000ffffffff  256M  online       yes    15         Normal
+0x0000000100000000-0x000000010fffffff  256M  online        no    16         Normal
+0x0000000110000000-0x000000011fffffff  256M  online        no    17 Normal/Movable
+0x0000000120000000-0x000000012fffffff  256M  online       yes    18 Movable/Normal
+0x0000000130000000-0x000000013fffffff  256M  online       yes    19        Movable
+0x0000000140000000-0x000000017fffffff    1G offline         - 20-23        Movable
+
+Memory block size:       256M
+Total online memory:     4.8G
+Total offline memory:    1.3G

index 14d7d84f65ca..40316a584518 100644
@@ -269,3 +269,42 @@ $ lsmem --json --output RANGE,SIZE,STATE,REMOVABLE,BLOCK,NODE
       {"range": "0x0000000438000000-0x000000043fffffff", "size": "128M", "state": "online", "removable": "no", "block": "135", "node": "0"}
    ]
 }
+
+---
+
+$ lsmem -o +ZONES
+RANGE                                  SIZE  STATE REMOVABLE   BLOCK  ZONES
+0x0000000000000000-0x0000000007ffffff  128M online        no       0   None
+0x0000000008000000-0x0000000037ffffff  768M online       yes     1-6  DMA32
+0x0000000038000000-0x000000003fffffff  128M online        no       7  DMA32
+0x0000000040000000-0x0000000077ffffff  896M online       yes    8-14  DMA32
+0x0000000078000000-0x000000007fffffff  128M online        no      15  DMA32
+0x0000000080000000-0x00000000afffffff  768M online       yes   16-21  DMA32
+0x00000000b0000000-0x00000000bfffffff  256M online        no   22-23  DMA32
+0x0000000100000000-0x00000001a7ffffff  2.6G online        no   32-52 Normal
+0x00000001a8000000-0x00000001afffffff  128M online       yes      53 Normal
+0x00000001b0000000-0x00000001bfffffff  256M online        no   54-55 Normal
+0x00000001c0000000-0x00000001ffffffff    1G online       yes   56-63 Normal
+0x0000000200000000-0x0000000207ffffff  128M online        no      64 Normal
+0x0000000208000000-0x000000021fffffff  384M online       yes   65-67 Normal
+0x0000000220000000-0x0000000237ffffff  384M online        no   68-70 Normal
+0x0000000238000000-0x0000000277ffffff    1G online       yes   71-78 Normal
+0x0000000278000000-0x000000028fffffff  384M online        no   79-81 Normal
+0x0000000290000000-0x0000000297ffffff  128M online       yes      82 Normal
+0x0000000298000000-0x00000002a7ffffff  256M online        no   83-84 Normal
+0x00000002a8000000-0x00000002c7ffffff  512M online       yes   85-88 Normal
+0x00000002c8000000-0x00000002dfffffff  384M online        no   89-91 Normal
+0x00000002e0000000-0x00000002efffffff  256M online       yes   92-93 Normal
+0x00000002f0000000-0x000000034fffffff  1.5G online        no  94-105 Normal
+0x0000000350000000-0x0000000357ffffff  128M online       yes     106 Normal
+0x0000000358000000-0x000000036fffffff  384M online        no 107-109 Normal
+0x0000000370000000-0x0000000377ffffff  128M online       yes     110 Normal
+0x0000000378000000-0x00000003c7ffffff  1.3G online        no 111-120 Normal
+0x00000003c8000000-0x00000003e7ffffff  512M online       yes 121-124 Normal
+0x00000003e8000000-0x000000042fffffff  1.1G online        no 125-133 Normal
+0x0000000430000000-0x0000000437ffffff  128M online       yes     134 Normal
+0x0000000438000000-0x000000043fffffff  128M online        no     135   None
+
+Memory block size:       128M
+Total online memory:      16G
+Total offline memory:      0B

index 79c0523b96a3..b1313773e212 100755
@@ -49,6 +49,7 @@ for dump in $(ls $TS_SELF/dumps/*.tar.bz2 | sort); do
 	do_lsmem --all --output $LSCOLUMNS
 	do_lsmem --raw --output $LSCOLUMNS
 	do_lsmem --json --output $LSCOLUMNS
+	do_lsmem -o +ZONES
 
 	ts_finalize_subtest
 done

2.13.5

From: Andre Wild <[email protected]>

This patch extends the valid --output values with ZONES for the lsmem bash-completion, and adds the --zone option for the chmem bash-completion.

---
 bash-completion/chmem | 1 +
 bash-completion/lsmem | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

index 00b870dbd90d..3e3af87acaa5 100644
@@ -16,6 +16,7 @@ _chmem_module()
 				--disable
 				--blocks
 				--verbose
+				--zone
 				--help
 				--version
 			"

index 8f7a46ec30af..9aa124569d53 100644
@@ -9,7 +9,7 @@ _lsmem_module()
 			local prefix realcur OUTPUT_ALL OUTPUT
 			realcur="${cur##*,}"
 			prefix="${cur%$realcur}"
-			OUTPUT_ALL='RANGE SIZE STATE REMOVABLE BLOCK NODE'
+			OUTPUT_ALL='RANGE SIZE STATE REMOVABLE BLOCK NODE ZONES'
 			for WORD in $OUTPUT_ALL; do
 				if ! [[ $prefix == *"$WORD"* ]]; then
 					OUTPUT="$WORD ${OUTPUT:-""}"

2.13.5

Seems good.

I'll merge it (probably with some minor changes:-) after v2.31 release. Thanks!

Karel

 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com

Merged to my "next" branch (in master we have still feature-freeze).

I have also added a note about the way how lsmem merges blocks to create the RANGE column. It seems important, because the number of ranges is affected by ZONES (or REMOVABLE). See:

https://github.com/karelzak/util-linux/commit/ffe5267c91018ca8cac8bedc14b695478c11a5dd

Maybe it's possible to explain it in a better way... (send patch;-)

The another possibility is to *always use* zones and removable attributes to create the ranges (merge blocks) independently on the output columns (e.g. -o ZONES). So, the result will be always the same number of ranges with the same <start>-<end>.

Now (see the first range):

$ lsmem 
RANGE                                  SIZE  STATE REMOVABLE  BLOCK
0x0000000000000000-0x0000000047ffffff  1.1G online        no    0-8
0x0000000048000000-0x0000000057ffffff  256M online       yes   9-10
0x0000000058000000-0x000000005fffffff  128M online        no     11
0x0000000060000000-0x0000000067ffffff  128M online       yes     12
0x0000000068000000-0x0000000087ffffff  512M online        no  13-16
0x0000000088000000-0x000000008fffffff  128M online       yes     17
0x0000000090000000-0x00000000afffffff  512M online        no  18-21
0x00000000b0000000-0x00000000bfffffff  256M online       yes  22-23
0x0000000100000000-0x000000042fffffff 12.8G online        no 32-133
0x0000000430000000-0x0000000437ffffff  128M online       yes    134
0x0000000438000000-0x000000043fffffff  128M online        no    135

lsmem -o+ZONES
RANGE SIZE STATE REMOVABLE BLOCK ZONES
0x0000000000000000-0x0000000007ffffff 128M online no 0 None

0x0000000008000000-0x0000000047ffffff    1G online        no    1-8  DMA32
0x0000000048000000-0x0000000057ffffff  256M online       yes   9-10  DMA32
0x0000000058000000-0x000000005fffffff  128M online        no     11  DMA32
0x0000000060000000-0x0000000067ffffff  128M online       yes     12  DMA32
0x0000000068000000-0x0000000087ffffff  512M online        no  13-16  DMA32
0x0000000088000000-0x000000008fffffff  128M online       yes     17  DMA32
0x0000000090000000-0x00000000afffffff  512M online        no  18-21  DMA32
0x00000000b0000000-0x00000000bfffffff  256M online       yes  22-23  DMA32
0x0000000100000000-0x000000042fffffff 12.8G online        no 32-133 Normal
0x0000000430000000-0x0000000437ffffff  128M online       yes    134 Normal
0x0000000438000000-0x000000043fffffff  128M online        no    135   None

lsmem -oRANGE,SIZE
RANGE SIZE
0x0000000000000000-0x00000000bfffffff 3G
0x0000000100000000-0x000000043fffffff 13G

I didn't test it, but the question is how usable is 0x0000000000000000-<end> as option for chmem.

It's also seems difficult to use it in scripts if you want to output only a RANGE, for example

    FOO=$(lsmem -oRANGE -n --summary=never | head -1)

but the range is affected by missing columns.

Comments?

Karel

 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com

Sorry for the late answer. I'm not sure if I understand the problem, it "works as designed" that the range merging is done based on the output columns, but I see that it was not really described as such. So I do like the note that you added with the above mentioned commit.

However, regarding the --split option, I think it may be confusing at least for human users, if an "lsmem -oRANGE" will now print more than one range, even if this is now based on a "fixed" set of default columns that are used for merging (but "subject to change" according to the man page).

Now the user will not see the columns that are used for merging if he says "lsmem -oRANGE", as opposed to the previous behavior where only the visible output columns were used for merge decision (and for "lsmem -oRANGE" that would simply be one big range because there are no other columns that may differ).

I also do not really see the benefit for script usage, at least if we define it as "expected behavior" to have the ranges merged based on the output columns. Maybe I am missing something, but I think the --split option does not really solve any problem, but rather introduces potential for future confusion. I would rather only have the behavior documented in the man pages, as you did in the above mentioned commit (and which was then removed with the --split patch).

Regards,
Gerald

OK, I think we can support both concepts :-) I have modified lsmem to:

- follow output columns for split policy by default (= your original implementation)
- the --split is optional and may be used to override the default behavior

it means for humans it's probably less concussing and advanced users may define by --split another way how to generate the ranges.

I think it's good compromise and it's backwardly compatible with the previous version. OK?

If yes, I need to backport this change to RHEL7.5 :-)

We want to keep it user friendly. The "expected behavior" (now default) forces you to parse lsmem output to filter out unnecessary columns (if you care about RANGE only).

And in all our utils the --output option really control the output, but no another behavior.

Karel

 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com

Yes, that looks good.

Yes, please :-)

OK, that makes sense. I did not have any output selection in the original implementation, and also no focus on script usage, but as (maybe so far the only) human user I did get confused by the new range merging.

Regards,
Gerald

Merged to the master branch.

I have added a new --split=<list> command line option to the lsmem.

It allows to control the way how lsmem merges memory blocks to the ranges. The default is to use differences in STATE,REMOVABLE,NODE and ZONES attributes. The ranges are no more affected by --output, it means commands like

    lsmem --output=RANGE
    lsmem --output=RANGE,ZONES

returns the same memory ranges. You have to use --split to define your policy:

    lsmem --split=REMOVABLE,STATE

will create the ranges according to REMOVABLE and STATE attributes.

Karel

 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com