It can slowdown annotation browser if objdump is processing large DWARF
data. Let's add a hashmap to save the data type info for each line.
Note that this is needed for TUI only because stdio only processes each
line once. TUI will display the same line whenever it refreshes the
screen.
Signed-off-by: Namhyung Kim <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Ian Rogers <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
[ Add lines around an if block and use zfree() in one case, acked by Namhyung ]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
Recently it uses llvm and capstone to speed up annotation or disassembly
of instructions. But they don't support source code view yet. Until it
fixed, we can force to use objdump for source code annotation.
To prevent performance loss, it's disabled by default and turned it on
when user requests it in TUI by pressing 's' key.
Acked-by: Ian Rogers <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Reported-by: Ingo Molnar <[email protected]>
Signed-off-by: Namhyung Kim <[email protected]>
Sometimes it's useful to see both instructions and their data type
together. Let's extend the annotate code to use data type profiling
functions.
To make it easy to pass more argument, introduce a struct to carry
necessary information together. Also add a new annotation_option called
'code_with_type' to control the behavior. This is not enabled yet but
it'll be set later from the command line.
For simplicity, this is implemented for --stdio only.
Reviewed-by: Ian Rogers <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Namhyung Kim <[email protected]>
So that it can only handle a single disasm_linme and hopefully make the
code simpler. This is also a preparation to be called from different
places later.
The NO_TYPE macro was added to distinguish when it failed or needs retry.
Reviewed-by: Ian Rogers <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Namhyung Kim <[email protected]>
It's a prepartion to support code annotation and data type
annotation at the same time. Data type annotation needs more
information in the hist_entry so it needs to be passed deeper.
Also rename a function with the same name in the builtin-annotate.c
to hist_entry__stdio_annotate since it matches better to the command
line option. And change the condition inside to be simpler.
Reviewed-by: Ian Rogers <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Namhyung Kim <[email protected]>
The annotation_line__print() has many arguments. But min_percent,
max_lines and percent_type are from struct annotaion_options. So let's
pass a pointer to the option instead of passing them separately to
reduce the number of function arguments.
Actually it has a recursive call if 'queue' is set. Add a new option
instance to pass different values for the case.
Reviewed-by: Ian Rogers <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Namhyung Kim <[email protected]>
Prior to this change a string was used which could cause issues with
an unrecognized disassembler in symbol__disassembler. Change to
initializing an array of perf_disassembler enum values. If a value
already exists then adding it a second time is ignored to avoid array
out of bounds problems present in the previous code, it also allows a
statically sized array and removes memory allocation needs. Errors in
the disassembler string are reported when the config is parsed during
perf annotate or perf top start up. If the array is uninitialized
after processing the config file the default llvm, capstone then
objdump values are added but without a need to parse a string.
Fixes: a6e8a58de6 ("perf disasm: Allow configuring what disassemblers to use")
Closes: https://lore.kernel.org/lkml/CAP-5=fUdfCyxmEiTpzS2uumUp3-SyQOseX2xZo81-dQtWXj6vA@mail.gmail.com/
Signed-off-by: Ian Rogers <[email protected]>
Tested-by: Namhyung Kim <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Namhyung Kim <[email protected]>
The perf tools annotation code used for a long time parsing the output
of binutils's objdump (or its reimplementations, like llvm's) to then
parse and augment it with samples, allow navigation, etc.
More recently disassemblers from the capstone and llvm (libraries, not
parsing the output of tools using those libraries to mimic binutils's
objdump output) were introduced.
So when all those methods are available, there is a static preference
for a series of attempts of disassembling a binary, with the 'llvm,
capstone, objdump' sequence being hard coded.
This patch allows users to change that sequence, specifying via a 'perf
config' 'annotate.disassemblers' entry which and in what order
disassemblers should be attempted.
As alluded to in the comments in the source code of this series, this
flexibility is useful for users and developers alike, elliminating the
requirement to rebuild the tool with some specific set of libraries to
see how the output of disassembling would be for one of these methods.
root@x1:~# rm -f ~/.perfconfig
root@x1:~# perf annotate -v --stdio2 update_load_avg
<SNIP>
symbol__disassemble:
filename=/usr/lib/debug/lib/modules/6.11.4-201.fc40.x86_64/vmlinux,
sym=update_load_avg, start=0xffffffffb6148fe0, en>
annotating [0x6ff7170]
/usr/lib/debug/lib/modules/6.11.4-201.fc40.x86_64/vmlinux :
[0x7407ca0] update_load_avg
Disassembled with llvm
annotate.disassemblers=llvm,capstone,objdump
Samples: 66 of event 'cpu_atom/cycles/P', 10000 Hz,
Event count (approx.): 5185444, [percent: local period]
update_load_avg()
/usr/lib/debug/lib/modules/6.11.4-201.fc40.x86_64/vmlinux
Percent 0xffffffff81148fe0 <update_load_avg>:
1.61 pushq %r15
pushq %r14
1.00 pushq %r13
movl %edx,%r13d
1.90 pushq %r12
pushq %rbp
movq %rsi,%rbp
pushq %rbx
movq %rdi,%rbx
subq $0x18,%rsp
15.14 movl 0x1a4(%rdi),%eax
root@x1:~# perf config annotate.disassemblers=capstone
root@x1:~# cat ~/.perfconfig
# this file is auto-generated.
[annotate]
disassemblers = capstone
root@x1:~#
root@x1:~# perf annotate -v --stdio2 update_load_avg
<SNIP>
Disassembled with capstone
annotate.disassemblers=capstone
Samples: 66 of event 'cpu_atom/cycles/P', 10000 Hz,
Event count (approx.): 5185444, [percent: local period]
update_load_avg()
/usr/lib/debug/lib/modules/6.11.4-201.fc40.x86_64/vmlinux
Percent 0xffffffff81148fe0 <update_load_avg>:
1.61 pushq %r15
pushq %r14
1.00 pushq %r13
movl %edx,%r13d
1.90 pushq %r12
pushq %rbp
movq %rsi,%rbp
pushq %rbx
movq %rdi,%rbx
subq $0x18,%rsp
15.14 movl 0x1a4(%rdi),%eax
root@x1:~# perf config annotate.disassemblers=objdump,capstone
root@x1:~# perf config annotate.disassemblers
annotate.disassemblers=objdump,capstone
root@x1:~# cat ~/.perfconfig
# this file is auto-generated.
[annotate]
disassemblers = objdump,capstone
root@x1:~# perf annotate -v --stdio2 update_load_avg
Executing: objdump --start-address=0xffffffff81148fe0 \
--stop-address=0xffffffff811497aa \
-d --no-show-raw-insn -S -C "$1"
Disassembled with objdump
annotate.disassemblers=objdump,capstone
Samples: 66 of event 'cpu_atom/cycles/P', 10000 Hz,
Event count (approx.): 5185444, [percent: local period]
update_load_avg()
/usr/lib/debug/lib/modules/6.11.4-201.fc40.x86_64/vmlinux
Percent
Disassembly of section .text:
ffffffff81148fe0 <update_load_avg>:
#define DO_ATTACH 0x4
ffffffff81148fe0 <update_load_avg>:
#define DO_ATTACH 0x4
#define DO_DETACH 0x8
/* Update task and its cfs_rq load average */
static inline void update_load_avg(struct cfs_rq *cfs_rq,
struct sched_entity *se,
int flags)
{
1.61 push %r15
push %r14
1.00 push %r13
mov %edx,%r13d
1.90 push %r12
push %rbp
mov %rsi,%rbp
push %rbx
mov %rdi,%rbx
sub $0x18,%rsp
}
/* rq->task_clock normalized against any time
this cfs_rq has spent throttled */
static inline u64 cfs_rq_clock_pelt(struct cfs_rq *cfs_rq)
{
if (unlikely(cfs_rq->throttle_count))
15.14 mov 0x1a4(%rdi),%eax
root@x1:~#
After adding a way to select the disassembler from the command line a
'perf test' comparing the output of the various diassemblers should be
introduced, to test these codebases.
Acked-by: Ian Rogers <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Athira Rajeev <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Steinar H. Gunderson <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
No event is printed in the "Branch Counter" column on hybrid machines.
For example,
$ perf record -e "{cpu_core/branch-instructions/pp,cpu_core/branches/}:S" -j any,counter
$ perf report --total-cycles
# Branch counter abbr list:
# cpu_core/branch-instructions/pp = A
# cpu_core/branches/ = B
# '-' No event occurs
# '+' Event occurrences may be lost due to branch counter saturated
#
# Sampled Cycles% Sampled Cycles Avg Cycles% Avg Cycles Branch Counter
# ............... .............. ........... .......... ..............
44.54% 727.1K 0.00% 1 |+ |+ |
36.31% 592.7K 0.00% 2 |+ |+ |
17.83% 291.1K 0.00% 1 |+ |+ |
The branch counter information (br_cntr_width and br_cntr_nr) in the
perf_env is retrieved from the CPU_PMU_CAPS. However, the CPU_PMU_CAPS
is not available on hybrid machines. Without the width information, the
number of occurrences of an event cannot be calculated.
For a hybrid machine, the caps information should be retrieved from the
PMU_CAPS, and stored in the perf_env->pmu_caps.
Add a perf_env__find_br_cntr_info() to return the correct branch counter
information from the corresponding fields.
Committer notes:
While testing I couldn't s ee those "Branch counter" columns enabled by
pressing 'B' on the TUI, after reporting it to the list Kan explained
the situation:
<quote Kan Liang>
For a hybrid client, the "Branch Counter" feature is only supported
starting from the just released Lunar Lake. Perf falls back to only
"ANY" on your Raptor Lake.
The "The branch counter is not available" message is expected.
Here is the 'perf evlist' result from my Lunar Lake machine,
# perf evlist -v
cpu_core/branch-instructions/pp: type: 4 (cpu_core), size: 136, config: 0xc4 (branch-instructions), { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|READ|PERIOD|BRANCH_STACK|IDENTIFIER, read_format: ID|GROUP|LOST, disabled: 1, freq: 1, enable_on_exec: 1, precise_ip: 2, sample_id_all: 1, exclude_guest: 1, branch_sample_type: ANY|COUNTERS
#
</quote>
Fixes: 6f9d8d1de2 ("perf script: Add branch counters")
Reviewed-by: Ian Rogers <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
When annotating a basic block, it's useful to display the occurrences
of other events in the block.
The branch counter feature is only available for newer Intel platforms.
So a dedicated option to display the branch counters is not introduced.
Reuse the existing --total-cycles option, which triggers the annotation
of a basic block and displays the cycle-related annotation.
When the branch counters information is available, the branch counters
are automatically appended after all the cycle-related annotation.
Accounting the branch counters as well when accounting the cycles in
hist__account_cycles().
In 'struct annotated_branch', introduce a br_cntr array to save the
accumulation of each branch counter.
In a sample, all the branch counters for a branch are saved in a u64
space.
Because the saturation of a branch counter is small, e.g., for Intel
Sierra Forest, the saturation is only 3.
Add ANNOTATION__BR_CNTR_SATURATED_FLAG to indicate if a branch counter
once saturated. That can be used to indicate a potential event lost
because of the saturation.
Reviewed-by: Andi Kleen <[email protected]>
Signed-off-by: Kan Liang <[email protected]>
Acked-by: Namhyung Kim <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Ian Rogers <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Stephane Eranian <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
In find_data_type(), it creates and deletes a debug info whenver it
tries to find data type for a sample. This is inefficient and it most
likely accesses the same binary again and again.
Let's add a single entry cache the debug info structure for the last DSO.
Depending on sample data, it usually gives me 2~3x (and sometimes more)
speed ups.
Note that this will introduce a little difference in the output due to
the order of checking stack operations. It used to check the stack ops
before checking the availability of debug info but I moved it after the
symbol check. So it'll report stack operations in DSOs without debug
info as unknown. But I think it's ok and better to have the checking
near the caching logic.
Committer testing:
root@x1:~# perf mem record -a sleep 5s
root@x1:~# perf evlist
cpu_atom/mem-loads,ldlat=30/P
cpu_atom/mem-stores/P
dummy:u
root@x1:~# diff -u before after
--- before 2024-08-08 09:33:53.880780784 -0300
+++ after 2024-08-08 09:35:13.917325041 -0300
@@ -81,8 +81,8 @@
# Overhead Data Type
# ........ .........
#
- 55.43% (unknown)
- 11.61% (stack operation)
+ 55.56% (unknown)
+ 11.48% (stack operation)
4.93% struct pcpu_hot
3.26% unsigned int
2.48% struct
Signed-off-by: Namhyung Kim <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Athira Rajeev <[email protected]>
Cc: Ian Rogers <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
The annotation__pcnt_width() calculates the screen width for the
overhead (percent) area considering event groups properly. Use this
function consistently so that we can make sure it has similar output
in different modes. But there's a difference in stdio and tui output:
stdio uses 8 and tui uses 7 for a percent.
Let's use 8 and adjust the print width in __annotation_line__write()
properly.
Signed-off-by: Namhyung Kim <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Ian Rogers <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
Since the "ins.name" is not set while using raw instruction,
'perf annotate' with insn-stat gives wrong data:
Result from "./perf annotate --data-type --insn-stat":
Annotate Instruction stats
total 615, ok 419 (68.1%), bad 196 (31.9%)
Name : Good Bad
-----------------------------------------------------------
: 419 196
This patch sets "dl->ins.name" in arch specific function
"check_ppc_insn" while initialising "struct disasm_line".
Also update "ins_find" function to pass "struct disasm_line" as a
parameter so as to set its name field in arch specific call.
With the patch changes:
Annotate Instruction stats
total 609, ok 446 (73.2%), bad 163 (26.8%)
Name/opcode : Good Bad
-----------------------------------------------------------
58 : 323 80
32 : 49 43
34 : 33 11
OP_31_XOP_LDX : 8 20
40 : 23 0
OP_31_XOP_LWARX : 5 1
OP_31_XOP_LWZX : 2 3
OP_31_XOP_LDARX : 3 0
33 : 0 2
OP_31_XOP_LBZX : 0 1
OP_31_XOP_LWAX : 0 1
OP_31_XOP_LHZX : 0 1
Reviewed-by: Kajol Jain <[email protected]>
Reviewed-by: Namhyung Kim <[email protected]>
Signed-off-by: Athira Rajeev <[email protected]>
Tested-by: Kajol Jain <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Akanksha J N <[email protected]>
Cc: Christophe Leroy <[email protected]>
Cc: Disha Goel <[email protected]>
Cc: Hari Bathini <[email protected]>
Cc: Ian Rogers <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Madhavan Srinivasan <[email protected]>
Cc: Segher Boessenkool <[email protected]>
Link: https://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
Use the raw instruction code and macros to identify memory instructions,
extract register fields and also offset.
The implementation addresses the D-form, X-form, DS-form instructions.
Adds "mem_ref" field to check whether source/target has memory
reference.
Add function "get_powerpc_regs" which will set these fields: reg1, reg2,
offset depending of where it is source or target ops.
Update "parse" callback for "struct ins_ops" to also pass "struct
disasm_line" as argument. This is needed in parse functions where opcode
is used to determine whether to set multi_regs and other fields
Reviewed-by: Kajol Jain <[email protected]>
Reviewed-by: Namhyung Kim <[email protected]>
Signed-off-by: Athira Rajeev <[email protected]>
Tested-by: Kajol Jain <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Akanksha J N <[email protected]>
Cc: Christophe Leroy <[email protected]>
Cc: Disha Goel <[email protected]>
Cc: Hari Bathini <[email protected]>
Cc: Ian Rogers <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Madhavan Srinivasan <[email protected]>
Cc: Segher Boessenkool <[email protected]>
Link: https://lore.kernel.org/lkml/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
When freeing a->b it is good practice to set a->b to NULL using
zfree(&a->b) so that when we have a bug where a reference to a freed 'a'
pointer is kept somewhere, we can more quickly cause a segfault if some
code tries to use a->b.
This is mostly done but some new cases were introduced recently, convert
them to zfree().
Cc: Adrian Hunter <[email protected]>
Cc: Ian Rogers <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Namhyung Kim <[email protected]>
Link: https://lore.kernel.org/lkml/ZjmbHHrjIm5YRIBv@x1
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
Add reference count checking to struct dso, this can help with
implementing correct reference counting discipline. To avoid
RC_CHK_ACCESS everywhere, add accessor functions for the variables in
struct dso.
The majority of the change is mechanical in nature and not easy to
split up.
Committer testing:
'perf test' up to this patch shows no regressions.
But:
util/symbol.c: In function ‘dso__load_bfd_symbols’:
util/symbol.c:1683:9: error: too few arguments to function ‘dso__set_adjust_symbols’
1683 | dso__set_adjust_symbols(dso);
| ^~~~~~~~~~~~~~~~~~~~~~~
In file included from util/symbol.c:21:
util/dso.h:268:20: note: declared here
268 | static inline void dso__set_adjust_symbols(struct dso *dso, bool val)
| ^~~~~~~~~~~~~~~~~~~~~~~
make[6]: *** [/home/acme/git/perf-tools-next/tools/build/Makefile.build:106: /tmp/tmp.ZWHbQftdN6/util/symbol.o] Error 1
MKDIR /tmp/tmp.ZWHbQftdN6/tests/workloads/
make[6]: *** Waiting for unfinished jobs....
This was updated:
- symbols__fixup_end(&dso->symbols, false);
- symbols__fixup_duplicate(&dso->symbols);
- dso->adjust_symbols = 1;
+ symbols__fixup_end(dso__symbols(dso), false);
+ symbols__fixup_duplicate(dso__symbols(dso));
+ dso__set_adjust_symbols(dso);
But not build tested with BUILD_NONDISTRO and libbfd devel files installed
(binutils-devel on fedora).
Add the missing argument:
symbols__fixup_end(dso__symbols(dso), false);
symbols__fixup_duplicate(dso__symbols(dso));
- dso__set_adjust_symbols(dso);
+ dso__set_adjust_symbols(dso, true);
Signed-off-by: Ian Rogers <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Ahelenia Ziemiańska <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: Athira Rajeev <[email protected]>
Cc: Ben Gainey <[email protected]>
Cc: Changbin Du <[email protected]>
Cc: Chengen Du <[email protected]>
Cc: Colin Ian King <[email protected]>
Cc: Dima Kogan <[email protected]>
Cc: Ilkka Koskinen <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: James Clark <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: K Prateek Nayak <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Leo Yan <[email protected]>
Cc: Li Dong <[email protected]>
Cc: Mark Rutland <[email protected]>
Cc: Masami Hiramatsu <[email protected]>
Cc: Namhyung Kim <[email protected]>
Cc: Paran Lee <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Song Liu <[email protected]>
Cc: Sun Haiyong <[email protected]>
Cc: Thomas Richter <[email protected]>
Cc: Tiezhu Yang <[email protected]>
Cc: Yanteng Si <[email protected]>
Cc: zhaimingbing <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
For data type profiling, it removed non-instruction lines from the list
of annotation lines. It was to simplify the implementation dealing with
instructions like to calculate the PC-relative address and to search the
shortest path to the target instruction or basic block.
But it means that it removes all the comments and debug information in
the annotate output like source file name and line numbers. To support
both code annotation and data type annotation, it'd be better to keep
the non-instruction lines as well.
So this change is to skip those lines during the data type profiling
and to display them in the normal perf annotate output.
No function changes intended (other than having more lines).
Reviewed-by: Ian Rogers <[email protected]>
Signed-off-by: Namhyung Kim <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
The symbol__annotate2() initializes some data structures needed by TUI.
It has a logic to prevent calling it multiple times by checking if it
has the annotated source. But data type profiling uses a different
code (symbol__annotate) to allocate the annotated lines in advance.
So TUI missed to call symbol__annotate2() when it shows the annotation
browser.
Make symbol__annotate() reentrant and handle that situation properly.
This fixes a crash in the annotation browser started by perf report in
TUI like below.
$ perf report -s type,sym --tui
# and press 'a' key and then move down
Fixes: 81e57deec3 ("perf report: Support data type profiling")
Reviewed-by: Ian Rogers <[email protected]>
Signed-off-by: Namhyung Kim <[email protected]>
Link: https://lore.kernel.org/r/[email protected]