[PATCH v7 00/13] Unify the interrupt delivery mode and do its setup in advance

From: Dou Liyang
Date: Fri Jul 14 2017 - 01:53:33 EST

[Background]

MP specification defines three different interrupt delivery modes as follows:

1. PIC Mode
2. Virtual Wire Mode
3. Symmetric I/O Mode

They will be setup in the different periods of booting time:
1. *PIC Mode*, the default interrupt delivery modes, will be set first.
2. *Virtual Wire Mode* will be setup during ISA IRQ initialization( step 1
in the figure.1).
3. *Symmetric I/O Mode*'s setup is related to the system
3.1 In SMP-capable system, setup during prepares CPUs(step 2)
3.2 In UP system, setup during initializes itself(step 3).


start_kernel
+---------------+
|
+--> .......
|
| setup_arch
+--> +-------+
|
| init_IRQ
+-> +--+-----+
| | init_ISA_irqs
| +------> +-+--------+
| | +----------------+
+---> +------> | 1.init_bsp_APIC|
| ....... +----------------+
+--->
| rest_init
+--->---+-----+
| | kernel_init
| +> ----+-----+
| | kernel_init_freeable
| +-> ----+-------------+
| | smp_prepare_cpus
| +---> +----+---------+
| | | +-------------------+
| | +-> |2. apic_bsp_setup |
| | +-------------------+
| |
v | smp_init
+---> +---+----+
| +-------------------+
+--> |3. apic_bsp_setup |
+-------------------+
figure.1 The flow chart of the kernel startup process

[Problem]

1. Cause kernel in an unmatched mode at the beginning of booting time.
2. Cause the dump-capture kernel hangs with 'notsc' option inherited
from 1st kernel option.
3. Cause the code hard to read and maintain.

As Ingo's and Eric's discusses[1,2], it need to be refactor.

[Solution]

1. Construct a selector to unify these switches

+------------+
|disable_apic+--------------------+
+------------+ true |
|false |
| |
+------------v------------------+ |
|!boot_cpu_has(X86_FEATURE_APIC)+-------+
+-------------------------------+ true |
|false |
| |
+-------v---------+ v
|!smp_found_config| PIC MODE
+---------------+-+
|false |true
| |
v +---v---------+
SYMMETRIC IO MODE | !acpi_lapic |
+------+------+
|
v
VIRTUAL WIRE MODE

2. Unifying these setup steps of SMP-capable and UP system

start_kernel
---------------+
|
|
|
| x86_late_time_init
+---->---+------------+
| |
| | +------------------------+
| +----> | 4. init_interrupt_mode |
| +------------------------+
v


3. Execute the function as soon as possible.

[Test]

1. In a theoretical code analysis, the patchset can wrap the original
logic.

1) The original logic of the interrupt delivery mode setupï

-Step O_1) Keep in PIC mode or virtual wire mode:

Check (smp_found_config || !boot_cpu_has(X86_FEATURE_APIC))
true: PIC mode
false: virtual wire mode

-Step O_2) Try to switch to symmetric IO mode:
O_2_1) In up system:

-Check disable_apic
ture: O_S_1 (original situation 1)
-Check whether there is a separate or integrated chip
don't has: O_S_2
-Check !smp_found_config
ture: O_S_3
-Others:
O_S_4

O_2_2) In smp-capable system:

-Check !smp_found_config && !acpi_lapic
true: goto O_2_1
-Check if it is LAPIC
don't has: O_S_5
-Check !max_cpus
true: O_S_6
-read_apic_id() != boot_cpu_physical_apicid
true: O_S_7
-Others:
O_S_8

2) After that patchset, the new logic:

-Step N_1) Skip step O_1 and try to switch to the final interrupt mode
-Check disable_apic
ture: N_S_1 (New situation 1)
-Check whether there is a separate or integrated chip
ture: N_S_2
-Check if (!smp_found_config)
ture: N_S_3
-Check !setup_max_cpus
ture: N_S_4
-Check read_apic_id() != boot_cpu_physical_apicid
ture: N_S_5
-Others:
N_S_6

O_S_1 is covered in N_S_1
O_S_2 is covered in N_S_2
O_S_3 is covered in N_S_3
O_S_4 is covered in N_S_6
O_S_5 is covered in N_S_2
O_S_6 is covered in N_S_4
O_S_7 is covered in N_S_5
O_S_8 is covered in N_S_6

2. In the actual test, It also can work well in the situations of
my test matrix

The factors of test matrix:

X86 | SMP |LOCAL APIC|I/O APIC|UP_LATE_INIT|
----- |-----|----------|--------|------------|
32-bit| Y | Y | Y | Y |
64-bit| N | N | N | N |
xen PV |
xen HVM|

disable_apic|X86_FEATURE_APIC|smp_found_config|
------------|----------------|----------------|
0 | 0 | 0 |
1 | 1 | 1 |

acpi_lapic|acpi_ioapic|setup_max_cpus|
----------|-----------|--------------|
0 | 0 | =0 |
1 | 1 | >0 |

[Link]

[1]. https://lkml.org/lkml/2016/8/2/929
[2]. https://lkml.org/lkml/2016/8/1/506

For previous discussion, please refer to:
https://lkml.org/lkml/2017/5/10/323
https://www.spinics.net/lists/kernel/msg2491620.html
https://lkml.org/lkml/2017/3/29/481
https://lkml.org/lkml/2017/6/30/17
https://lkml.org/lkml/2017/7/3/269

Changes V6 --> V7:

- add a new patch: p12

Changes V5 --> V6:

- change the check order for X86_32 in apic_intr_mode_select()
- replace the apic_printk with pr_info in apic_intr_mode_init()
- add a seperate helper function for get the logical apicid
- remove the extra argument upmode in apic_intr_mode_select()
- cleanup the logic of apic_intr_mode_init()
- replcae the 'ticks = jiffies' with 'end = jiffies + 4'
- rewrite the 9th and 10th patches's changelog

Changes V4 --> V5:

- remove the RFC presix
- remove the 1/12 patch in V4
- merge 2 patches together for SMP-capable system
- replace the *_interrupt_* with *_intr_*
- replace the pr_info with apic_printk in apic_intr_mode_init()
- add a patch for PV xen to bypass intr_mode_init()

Changes V3 --> V4:

- Move interrupt_mode_init to x86_init_ops instead of the use of
header files
- Replace "return" with "break" in case of APIC_SYMMETRIC_IO_NO_ROUTING
- Setup upmode earlier for UP system.
- Check interrupt mode before per cpu clock event setup.

Changes V2 --> V3:

- Rebase the patches.
- Change two function name:
apic_bsp_mode_check --> apic_interrupt_mode_select
init_interrupt_mode --> apic_interrupt_mode_init
- Find a new waiting way to check whether timer IRQs work or not
- Refine the switch logic in apic_interrupt_mode_init()
- Consistently start sentences with upper case letters
- Fix some typos and comments
- Try my best to rewrite some changelog again

Changes since V1:

- Move the initialization from init_IRQ() to x86_late_time_init()
- Use a threshold to refactor the check logic in timer_irq_works()
- Rename the framework to a selector
- Split two patches
- Consistently start sentences with upper case letters
- Fix some typos
- Rewrite the changelog


Dou Liyang (13):
x86/apic: Construct a selector for the interrupt delivery mode
x86/apic: Prepare for unifying the interrupt delivery modes setup
x86/apic: Split local APIC timer setup from the APIC setup
x86/apic: Move logical APIC ID away from apic_bsp_setup()
x86/apic: Unify interrupt mode setup for SMP-capable system
x86/apic: Mark the apic_intr_mode extern for sanity check cleanup
x86/apic: Unify interrupt mode setup for UP system
x86/ioapic: Refactor the delay logic in timer_irq_works()
x86/init: add intr_mode_init to x86_init_ops
x86/xen: Bypass intr mode setup in enlighten_pv system
x86/time: Initialize interrupt mode behind timer init
ACPI / init: Invoke early ACPI initialization earlier
x86/apic: Remove the init_bsp_APIC()

arch/x86/include/asm/apic.h | 15 +++-
arch/x86/include/asm/x86_init.h | 2 +
arch/x86/kernel/apic/apic.c | 190 +++++++++++++++++++++-------------------
arch/x86/kernel/apic/io_apic.c | 45 +++++++++-
arch/x86/kernel/irqinit.c | 3 -
arch/x86/kernel/smpboot.c | 80 +++++------------
arch/x86/kernel/time.c | 5 ++
arch/x86/kernel/x86_init.c | 1 +
arch/x86/xen/enlighten_pv.c | 1 +
init/main.c | 2 +-
10 files changed, 188 insertions(+), 156 deletions(-)

--
2.5.5