x86: fix race in create_irq_nr on irq_desc
Race in create_irq_nr():
- Thread 1 loops through and calls irq_to_desc_alloc_node with new=0x66.
- Thread 2 has exited the loop with irq=0x66 and calls dynamic_irq_init(0x66) setting desc->chip_data = NULL
- Thread 1 then dereferences NULL via desc_new->chip_data->vector
Fix by moving holding vector_lock until after the dynamic_irq_init().
BUG: unable to handle kernel NULL pointer dereference at 0000000000000088 IP: [<ffffffff8101df32>] create_irq_nr+0x62/0x100 PGD 23dc24067 PUD 23dc72067 PMD 0 Oops: 0000 [#1] SMP last sysfs file: /sys/devices/pci0000:00/0000:00:1c.0/0000:08:00.0/net/eth2/type CPU 12 Modules linked in: i2c_i801 igb(+) iTCO_wdt ixgbe(+) ioatdma(+) e1000e mptctl mdio usb_storage iTCO_vendor_support dca ses button sg pcspkr enclosure container ac usbhid uhci_hcd ehci_hcd usbcore sd_mod edd fan processor ide_pci_generic ide_core ata_generic ata_piix libata lpfc scsi_transport_fc scsi_tgt mptsas mptscsih mptbase scsi_transport_sas megaraid_sas scsi_mod thermal thermal_sys Supported: Yes Pid: 1684, comm: modprobe Not tainted 2.6.32.3-0.3-default #1 PRIMERGY RX300 S5 RIP: 0010:[<ffffffff8101df32>] [<ffffffff8101df32>] create_irq_nr+0x62/0x100 RSP: 0018:ffff88013ce0fc18 EFLAGS: 00010086 RAX: ffff88023e11ee00 RBX: 0000000000000066 RCX: 00000000000000c2 RDX: 00000000000000c2 RSI: 00000000ffffffff RDI: 0000000000000066 RBP: 0000000000000000 R08: ffffffff81767a85 R09: 000000000000000a R10: 0000000000000000 R11: 0000000000000000 R12: 00000000ffffffff R13: 0000000000000206 R14: ffff88013f381000 R15: 0000000000000080 FS: 00007f16d181e700(0000) GS:ffff880143d00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: 0000000000000088 CR3: 000000023d26c000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Process modprobe (pid: 1684, threadinfo ffff88013ce0e000, task ffff88013d080340) Stack: 0000000000000001 0000000000000000 ffff88023d2d8740 0000000000000064 <0> 0000000000000007 ffffffff8101f2ce 0000000900000009 ffff88013f381810 <0> ffffffff3f381000 0000000000000048 0000000000000009 ffff88013f381000 Call Trace: [<ffffffff8101f2ce>] arch_setup_msi_irqs+0xce/0x190 [<ffffffff812574b9>] msix_capability_init+0x189/0x2f0 [<ffffffffa032b4a4>] igb_set_interrupt_capability+0xe4/0x1e0 [igb] [<ffffffffa033634e>] igb_probe+0x3de/0xd15 [igb] [<ffffffff8124d212>] local_pci_probe+0x12/0x20 [<ffffffff8124d4c0>] __pci_device_probe+0xe0/0xf0 [<ffffffff8124e3d3>] pci_device_probe+0x33/0x60 [<ffffffff812e72f7>] really_probe+0x77/0x230 [<ffffffff812e751a>] driver_probe_device+0x6a/0xc0 [<ffffffff812e7603>] __driver_attach+0x93/0xa0 [<ffffffff812e6928>] bus_for_each_dev+0x58/0x80 [<ffffffff812e6115>] bus_add_driver+0x195/0x2f0 [<ffffffff812e7919>] driver_register+0x79/0x170 [<ffffffff8124e648>] __pci_register_driver+0x58/0xe0 [<ffffffff810001e5>] do_one_initcall+0x35/0x190 [<ffffffff8108af34>] sys_init_module+0xe4/0x270 [<ffffffff81002f7b>] system_call_fastpath+0x16/0x1b [<00007f16d13b234a>] 0x7f16d13b234a Code: 2e 0f 1f 84 00 00 00 00 00 83 c3 01 39 1d e7 e2 9f 00 76 7d 44 89 e6 89 df e8 2b 2a 3d 00 48 85 c0 0f 84 8a 00 00 00 48 8b 68 40 <80> bd 88 00 00 00 00 75 d5 44 89 e6 48 89 c7 e8 6a 5c 09 00 49 RIP [<ffffffff8101df32>] create_irq_nr+0x62/0x100 RSP <ffff88013ce0fc18>
CR2: 0000000000000088
--- arch/x86/kernel/apic/io_apic.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
arch/x86/kernel/apic/io_apic.c +2 -1
@@ -3212,7 +3212,6 @@ unsigned int create_irq_nr(unsigned int irq = new;break ; } -spin_unlock_irqrestore (&vector_lock, flags);if (irq >0 ) {dynamic_irq_init (irq);
@@ -3220,6 +3219,8 @@ unsigned int create_irq_nr(unsigned intif (desc_new) desc_new->chip_data = cfg_new; } +spin_unlock_irqrestore (&vector_lock, flags); +return irq; }
Quoted 8 lines · Brandon Philips
Race in
create_irq_nr():- Thread 1 loops through and calls
irq_to_desc_alloc_nodewith new=0x66.- Thread 2 has exited the loop with irq=0x66 and calls
dynamic_irq_init(0x66)settingdesc->chip_data= NULL- Thread 1 then dereferences NULL via
desc_new->chip_data->vector
two threads get same irq?
YH
Quoted 11 lines · Yinghai Lu
Quoted 8 lines · Brandon Philips
Race in
create_irq_nr():- Thread 1 loops through and calls
irq_to_desc_alloc_nodewith new=0x66.- Thread 2 has exited the loop with irq=0x66 and calls
dynamic_irq_init(0x66)settingdesc->chip_data= NULL- Thread 1 then dereferences NULL via
desc_new->chip_data->vectortwo threads get same irq?
This race happened when two drivers were setting up MSI-X at the same time via pci_enable_msix(). See this dmesg excerpt:
[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereference
at 0000000000000088
As you can see igb and ixgbe are both alternating on create_irq_nr() via pci_enable_msix() in their probe function. So, let me rewrite my explanation using this example:
ixgbe: While looping through irq_desc_ptrs[] via create_irq_nr() ixgbe choses irq_desc_ptrs[102] and exits the loop, drops vector_lock and calls dynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL via dynamic_irq_init().
igb: Grabs the vector_lock now and starts looping over irq_desc_ptrs[] via create_irq_nr(). It gets to irq_desc_ptrs[102] and does this:
cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
This hits the NULL deref.
Does that make sense? It is sort of a rare thing to reproduce- took 40+ reboots.
Thanks,
Brandon
Quoted 44 lines · Brandon Philips
Quoted 11 lines · Yinghai Lu
Quoted 8 lines · Brandon Philips
Race in
create_irq_nr():- Thread 1 loops through and calls
irq_to_desc_alloc_nodewith new=0x66.- Thread 2 has exited the loop with irq=0x66 and calls
dynamic_irq_init(0x66)settingdesc->chip_data= NULL- Thread 1 then dereferences NULL via
desc_new->chip_data->vectortwo threads get same irq?
This race happened when two drivers were setting up MSI-X at the same time via
pci_enable_msix(). See this dmesg excerpt:[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereferenceat 0000000000000088
As you can see igb and ixgbe are both alternating on
create_irq_nr()viapci_enable_msix()in their probe function. So, let me rewrite my explanation using this example:ixgbe: While looping through irq_desc_ptrs[] via
create_irq_nr()ixgbe choses irq_desc_ptrs[102] and exits the loop, dropsvector_lockand callsdynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL viadynamic_irq_init().
Quoted 9 lines
igb: Grabs the
vector_locknow and starts looping over irq_desc_ptrs[] viacreate_irq_nr(). It gets to irq_desc_ptrs[102] and does this:cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;This hits the NULL deref.
please try following patch in addition to
arch/x86/kernel/apic/io_apic.c +3 -6
@@ -3280,12 +3280,9 @@ unsigned int create_irq_nr(unsigned int irq_want, int node) }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
include/linux/irq.h +1 -0
@@ -400,6 +400,7 @@ static inline int irq_has_action(unsigned int irq)/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq);/* Set/get chip/data for an IRQ: */
kernel/irq/chip.c +13 -2
@@ -22,7 +22,7 @@* dynamic_irq_init - initialize a dynamically allocated irq * @irq: irq number to initialize */ -void dynamic_irq_init (unsigned int irq) +static void dynamic_irq_init_x (unsigned int irq,bool keep_chip_data) {struct irq_desc *desc;unsigned long flags;
@@ -41,7 +41,8 @@ void dynamic_irq_init(unsigned int irq) desc->depth =1 ; desc->msi_desc =NULL ; desc->handler_data =NULL ; - desc->chip_data =NULL ; +if (!keep_chip_data) + desc->chip_data =NULL ; desc->action =NULL ; desc->irq_count =0 ; desc->irqs_unhandled =0 ;
@@ -54,6 +55,16 @@ void dynamic_irq_init(unsigned int irq)raw_spin_unlock_irqrestore (&desc->lock, flags); } +void dynamic_irq_init (unsigned int irq) +{ +dynamic_irq_init_x (irq, false); +} + +void dynamic_irq_init_keep_chip_data (unsigned int irq) +{ +dynamic_irq_init_x (irq, true); +} +/** * dynamic_irq_cleanup - cleanup a dynamically allocated irq * @irq: irq number to initialize
Quoted 52 lines · Yinghai Lu
Quoted 51 lines · Brandon Philips
Quoted 11 lines · Yinghai Lu
Quoted 8 lines · Brandon Philips
Race in
create_irq_nr():- Thread 1 loops through and calls
irq_to_desc_alloc_nodewith new=0x66.- Thread 2 has exited the loop with irq=0x66 and calls
dynamic_irq_init(0x66)settingdesc->chip_data= NULL- Thread 1 then dereferences NULL via
desc_new->chip_data->vectortwo threads get same irq?
This race happened when two drivers were setting up MSI-X at the same time via
pci_enable_msix(). See this dmesg excerpt:[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereferenceat 0000000000000088
As you can see igb and ixgbe are both alternating on
create_irq_nr()viapci_enable_msix()in their probe function. So, let me rewrite my explanation using this example:ixgbe: While looping through irq_desc_ptrs[] via
create_irq_nr()ixgbe choses irq_desc_ptrs[102] and exits the loop, dropsvector_lockand callsdynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL viadynamic_irq_init().igb: Grabs the
vector_locknow and starts looping over irq_desc_ptrs[] viacreate_irq_nr(). It gets to irq_desc_ptrs[102] and does this:cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
Quoted 6 lines
Quoted 2 lines
This hits the NULL deref.
please try following patch in addition to
How is this commit related to this bug? The NULL deref I am hitting is from this bit in create_irq_nr():
if (cfg_new->vector != 0)
continue;Which comes before the assignment of cfg_new. I don't see how it is related. Plus, node == -1 in this case so move_irq_desc() is a no-op.
Quoted 9 lines
arch/x86/kernel/apic/io_apic.c +0
-1@@ -3280,12 +3280,9 @@ unsigned int create_irq_nr(unsigned int irq_want, int node) }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) {
arch/x86/kernel/apic/io_apic.c unsigned int create_irq_nr(unsigned int irq_want, int node) · line 3287
Quoted 10 lines
arch/x86/kernel/apic/io_apic.c +3
-5-dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
That would solve it too but I don't think it is a great solution. Keeping the vector_lock until we are completely done setting up the irq is more straightforward and won't cost much time at all.
I am hesitant to have it tested since it is a really small race window, reproducing took 40+ reboots initially and looks technically correct.
Thanks,
Brandon
PATCH 1/1 x86: keep chip_data in create_irq_nr
In reply to Brandon PhilipsBrodon found:
race happened when two drivers were setting up MSI-X at the same time via pci_enable_msix(). See this dmesg excerpt:
[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereference
at 0000000000000088
As you can see igb and ixgbe are both alternating on create_irq_nr() via pci_enable_msix() in their probe function.
ixgbe: While looping through irq_desc_ptrs[] via create_irq_nr() ixgbe choses irq_desc_ptrs[102] and exits the loop, drops vector_lock and calls dynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL via dynamic_irq_init().
igb: Grabs the vector_lock now and starts looping over irq_desc_ptrs[] via create_irq_nr(). It gets to irq_desc_ptrs[102] and does this:
cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
This hits the NULL deref.
so let remove the save and restore code.
just don't clear it in that path
--- arch/x86/kernel/apic/io_apic.c | 9 +++------ include/linux/irq.h | 1 + kernel/irq/chip.c | 15 +++++++++++++-- 3 files changed, 17 insertions(+), 8 deletions(-)
arch/x86/kernel/apic/io_apic.c +3 -6
@@ -3280,12 +3280,9 @@ unsigned int create_irq_nr(unsigned int }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
include/linux/irq.h +1 -0
@@ -400,6 +400,7 @@ static inline int irq_has_action(unsigne/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq);/* Set/get chip/data for an IRQ: */
kernel/irq/chip.c +13 -2
@@ -22,7 +22,7 @@* dynamic_irq_init - initialize a dynamically allocated irq * @irq: irq number to initialize */ -void dynamic_irq_init (unsigned int irq) +static void dynamic_irq_init_x (unsigned int irq,bool keep_chip_data) {struct irq_desc *desc;unsigned long flags;
@@ -41,7 +41,8 @@ void dynamic_irq_init(unsigned int irq) desc->depth =1 ; desc->msi_desc =NULL ; desc->handler_data =NULL ; - desc->chip_data =NULL ; +if (!keep_chip_data) + desc->chip_data =NULL ; desc->action =NULL ; desc->irq_count =0 ; desc->irqs_unhandled =0 ;
@@ -54,6 +55,16 @@ void dynamic_irq_init(unsigned int irq)raw_spin_unlock_irqrestore (&desc->lock, flags); } +void dynamic_irq_init (unsigned int irq) +{ +dynamic_irq_init_x (irq, false); +} + +void dynamic_irq_init_keep_chip_data (unsigned int irq) +{ +dynamic_irq_init_x (irq, true); +} +/** * dynamic_irq_cleanup - cleanup a dynamically allocated irq * @irq: irq number to initialize
Re: [PATCH] x86: keep chip_data in create_irq_nr
Quoted 52 lines · Yinghai Lu
Brodon found:
race happened when two drivers were setting up MSI-X at the same time viapci_enable_msix(). See this dmesg excerpt:[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereferenceat 0000000000000088
As you can see igb and ixgbe are both alternating on
create_irq_nr()viapci_enable_msix()in their probe function.ixgbe: While looping through irq_desc_ptrs[] via
create_irq_nr()ixgbe choses irq_desc_ptrs[102] and exits the loop, dropsvector_lockand callsdynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL viadynamic_irq_init().igb: Grabs the
vector_locknow and starts looping over irq_desc_ptrs[] viacreate_irq_nr(). It gets to irq_desc_ptrs[102] and does this:cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;This hits the NULL deref.
so let remove the save and restore code.
just don't clear it in that pathIndex:
linux-2.6/arch/x86/kernel/apic/io_apic.c
===================================================================arch/x86/kernel/apic/io_apic.c +0
-1@@ -3280,12 +3280,9 @@ unsigned int create_irq_nr(unsigned int }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) {
arch/x86/kernel/apic/io_apic.c unsigned int create_irq_nr(unsigned int · line 3287
Quoted 10 lines
arch/x86/kernel/apic/io_apic.c +3
-5-dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
Nearly every function in kernel/irq/chip.c takes the desc->lock when manipulating the fields of the irq_desc including chip_data. Should create_irq_nr() do the same when getting the chip_data field?
I am just a bit confused on what protects the chip_data field now.
Actually, while looking at your patch there is a related race in destroy_irq() that I just noticed. This race could happen via pci_disable_msix() in a driver or in the number of error paths that call free_msi_irqs():
destroy_irq()dynamic_irq_cleanup() which sets desc->chip_data = NULL ...race window...desc->chip_data = cfg;
It could race with create_irq_nr() in the same way in the irq destroy path.
So, I will reply after this with a combined patch fixing this potential race along with the minor things below.
Cheers,
Brandon
include/linux/irq.h static inline int irq_has_action(unsigne · line 404
Quoted 11 lines
Index:
linux-2.6/include/linux/irq.h
===================================================================include/linux/irq.h +1
-0@@ -400,6 +400,7 @@ static inline int irq_has_action(unsigne/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq);
Missing extern?
kernel/irq/chip.c line 23
Quoted 8 lines
include/linux/irq.h +0
-0/* Set/get chip/data for an IRQ: */ Index:
linux-2.6/kernel/irq/chip.c
===================================================================kernel/irq/chip.c +0
-0@@ -22,7 +22,7 @@* dynamic_irq_init - initialize a dynamically allocated irq * @irq: irq number to initialize
Update kerndoc?
Quoted 31 lines
kernel/irq/chip.c +13
-1+static void dynamic_irq_init_x (unsigned int irq,bool keep_chip_data) {struct irq_desc *desc;unsigned long flags;@@ -41,7 +41,8 @@ void dynamic_irq_init(unsigned int irq) desc->depth =1 ; desc->msi_desc =NULL ; desc->handler_data =NULL ; - desc->chip_data =NULL ; +if (!keep_chip_data) + desc->chip_data =NULL ; desc->action =NULL ; desc->irq_count =0 ; desc->irqs_unhandled =0 ;@@ -54,6 +55,16 @@ void dynamic_irq_init(unsigned int irq)raw_spin_unlock_irqrestore (&desc->lock, flags); } +void dynamic_irq_init (unsigned int irq) +{ +dynamic_irq_init_x (irq, false); +} + +void dynamic_irq_init_keep_chip_data (unsigned int irq) +{ +dynamic_irq_init_x (irq, true); +} +/** * dynamic_irq_cleanup - cleanup a dynamically allocated irq * @irq: irq number to initialize
H. Peter Anvin
reviewerRe: [PATCH] x86: keep chip_data in create_irq_nr
Quoted 7 lines · Brandon Philips
Quoted 7 lines
Index:
linux-2.6/include/linux/irq.h
===================================================================include/linux/irq.h +0
-0@@ -400,6 +400,7 @@ static inline int irq_has_action(unsigne
Quoted 7 lines
include/linux/irq.h static inline int irq_has_action(unsigne · line 404
Quoted 4 lines
include/linux/irq.h +1
-0/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq);Missing extern?
Quite the opposite -- function prototypes don't need "extern", and putting "extern" on them is generally considered bad style.
-hpa
PATCH 1/1 x86: keep chip_data in create_irq_nr and destroy_irq
In reply to Yinghai LuWhen two drivers are setting up MSI-X at the same time via pci_enable_msix() there is a race. See this dmesg excerpt:
[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereference
at 0000000000000088
As you can see igb and ixgbe are both alternating on create_irq_nr() via pci_enable_msix() in their probe function.
ixgbe: While looping through irq_desc_ptrs[] via create_irq_nr() ixgbe choses irq_desc_ptrs[102] and exits the loop, drops vector_lock and calls dynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL via dynamic_irq_init().
igb: Grabs the vector_lock now and starts looping over irq_desc_ptrs[] via create_irq_nr(). It gets to irq_desc_ptrs[102] and does this:
cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
This hits the NULL deref.
Another possible race exists via pci_disable_msix() in a driver or in the number of error paths that call free_msi_irqs():
destroy_irq()dynamic_irq_cleanup() which sets desc->chip_data = NULL ...race window...desc->chip_data = cfg;
Remove the save and restore code for cfg in create_irq_nr() and destroy_irq() and take the desc->lock when checking the irq_cfg.
--- arch/x86/kernel/apic/io_apic.c | 14 +++-------- include/linux/irq.h | 2 + kernel/irq/chip.c | 52 +++++++++++++++++++++++++++++++++-------- 3 files changed, 49 insertions(+), 19 deletions(-)
arch/x86/kernel/apic/io_apic.c +4 -10
@@ -3228,12 +3228,9 @@ unsigned int create_irq_nr(unsigned int }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
@@ -3260,10 +3257,7 @@ void destroy_irq(unsigned int irq)/* store it, in case dynamic_irq_cleanup clear it */ desc =irq_to_desc (irq); - cfg = desc->chip_data; -dynamic_irq_cleanup (irq); -/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags);
include/linux/irq.h +2 -0
@@ -400,7 +400,9 @@ static inline int irq_has_action(unsigne/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +extern void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq); +extern void dynamic_irq_cleanup_keep_chip_data (unsigned int irq);/* Set/get chip/data for an IRQ: */ extern int set_irq_chip (unsigned int irq,struct irq_chip *chip);
kernel/irq/chip.c +43 -9
84 lines not on this page · raw message
Re: [PATCH] x86: keep chip_data in create_irq_nr and destroy_irq
Quoted 174 lines · Brandon Philips
When two drivers are setting up MSI-X at the same time via
pci_enable_msix()there is a race. See this dmesg excerpt:[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereferenceat 0000000000000088
As you can see igb and ixgbe are both alternating on
create_irq_nr()viapci_enable_msix()in their probe function.ixgbe: While looping through irq_desc_ptrs[] via
create_irq_nr()ixgbe choses irq_desc_ptrs[102] and exits the loop, dropsvector_lockand callsdynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL viadynamic_irq_init().igb: Grabs the
vector_locknow and starts looping over irq_desc_ptrs[] viacreate_irq_nr(). It gets to irq_desc_ptrs[102] and does this:cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;This hits the NULL deref.
Another possible race exists via
pci_disable_msix()in a driver or in the number of error paths that callfree_msi_irqs():
destroy_irq()dynamic_irq_cleanup()which setsdesc->chip_data= NULL ...race window...desc->chip_data= cfg;Remove the save and restore code for cfg in
create_irq_nr()anddestroy_irq()and take thedesc->lockwhen checking theirq_cfg.Reported-and-analyzed-by Brandon Philips <[email protected]>Signed-off-by Yinghai Lu <[email protected]>Signed-off-by Brandon Phiilps <[email protected]>--- arch/x86/kernel/apic/io_apic.c | 14 +++-------- include/linux/irq.h | 2 + kernel/irq/chip.c | 52 +++++++++++++++++++++++++++++++++-------- 3 files changed, 49 insertions(+), 19 deletions(-)Index:
linux-2.6/arch/x86/kernel/apic/io_apic.c
===================================================================arch/x86/kernel/apic/io_apic.c +4
-10@@ -3228,12 +3228,9 @@ unsigned int create_irq_nr(unsigned int }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }@@ -3260,10 +3257,7 @@ void destroy_irq(unsigned int irq)/* store it, in case dynamic_irq_cleanup clear it */ desc =irq_to_desc (irq); - cfg = desc->chip_data; -dynamic_irq_cleanup (irq); -/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags);Index:
linux-2.6/include/linux/irq.h
===================================================================include/linux/irq.h +2
-0@@ -400,7 +400,9 @@ static inline int irq_has_action(unsigne/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +extern void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq); +extern void dynamic_irq_cleanup_keep_chip_data (unsigned int irq);/* Set/get chip/data for an IRQ: */ extern int set_irq_chip (unsigned int irq,struct irq_chip *chip);Index:
linux-2.6/kernel/irq/chip.c
===================================================================kernel/irq/chip.c +23
-9@@ -18,11 +18,7 @@#include "internals.h" -/** -* dynamic_irq_init - initialize a dynamically allocated irq -* @irq: irq number to initialize -*/ -void dynamic_irq_init (unsigned int irq) +static void dynamic_irq_init_x (unsigned int irq,bool keep_chip_data) {struct irq_desc *desc;unsigned long flags;@@ -41,7 +37,8 @@ void dynamic_irq_init(unsigned int irq) desc->depth =1 ; desc->msi_desc =NULL ; desc->handler_data =NULL ; - desc->chip_data =NULL ; +if (!keep_chip_data) + desc->chip_data =NULL ; desc->action =NULL ; desc->irq_count =0 ; desc->irqs_unhandled =0 ;@@ -55,10 +52,26 @@ void dynamic_irq_init(unsigned int irq) }/** -* dynamic_irq_cleanup - cleanup a dynamically allocated irq +* dynamic_irq_init - initialize a dynamically allocated irq * @irq: irq number to initialize */ -void dynamic_irq_cleanup (unsigned int irq) +void dynamic_irq_init (unsigned int irq) +{ +dynamic_irq_init_x (irq, false); +} + +/** +* dynamic_irq_init_keep_chip_data - initialize a dynamically allocated irq +* @irq: irq number to initialize +* +* does not set irq_to_desc(irq)->chip_data to NULL +*/ +void dynamic_irq_init_keep_chip_data (unsigned int irq) +{ +dynamic_irq_init_x (irq, true); +} + +static void dynamic_irq_cleanup_x (unsigned int irq,bool keep_chip_data) {struct irq_desc *desc =irq_to_desc (irq);unsigned long flags;@@ -77,7 +90,8 @@ void dynamic_irq_cleanup(unsigned int ir } desc->msi_desc =NULL ; desc->handler_data =NULL ; - desc->chip_data =NULL ; +if (!keep_chip_data) + desc->chip_data =NULL ; desc->handle_irq = handle_bad_irq; desc->chip = &no_irq_chip; desc->name =NULL ;
kernel/irq/chip.c void dynamic_irq_cleanup(unsigned int ir · line 108
Quoted 10 lines
kernel/irq/chip.c +7
-0@@ -85,6 +99,26 @@ void dynamic_irq_cleanup(unsigned int irraw_spin_unlock_irqrestore (&desc->lock, flags); } +/** +* dynamic_irq_cleanup - cleanup a dynamically allocated irq +* @irq: irq number to initialize +*/ +void dynamic_irq_cleanup (unsigned int irq) +{ +dynamic_irq_init_x (irq, false);
should be dynamic_irq_cleanup_x here.
kernel/irq/chip.c void dynamic_irq_cleanup(unsigned int ir · line 119
Quoted 11 lines
kernel/irq/chip.c +11
-0+} + +/** +* dynamic_irq_cleanup_keep_chip_data - cleanup a dynamically allocated irq +* @irq: irq number to initialize +* +* does not set irq_to_desc(irq)->chip_data to NULL +*/ +void dynamic_irq_cleanup_keep_chip_data (unsigned int irq) +{ +dynamic_irq_init_x (irq, true);
should be dynamic_irq_cleanup_x
kernel/irq/chip.c void dynamic_irq_cleanup(unsigned int ir · line 124
Quoted 5 lines
kernel/irq/chip.c +2
-0+} +/** * set_irq_chip - set the irq chip for an irq
YH
Re: [PATCH] x86: keep chip_data in create_irq_nr and destroy_irq
Quoted 25 lines · Yinghai Lu
Quoted 11 lines · Brandon Philips
kernel/irq/chip.c +2
-1@@ -77,7 +90,8 @@ void dynamic_irq_cleanup(unsigned int ir } desc->msi_desc =NULL ; desc->handler_data =NULL ; - desc->chip_data =NULL ; +if (!keep_chip_data) + desc->chip_data =NULL ; desc->handle_irq = handle_bad_irq; desc->chip = &no_irq_chip; desc->name =NULL ;kernel/irq/chip.c void dynamic_irq_cleanup(unsigned int ir · line 108
Quoted 10 lines
kernel/irq/chip.c +7
-0@@ -85,6 +99,26 @@ void dynamic_irq_cleanup(unsigned int irraw_spin_unlock_irqrestore (&desc->lock, flags); } +/** +* dynamic_irq_cleanup - cleanup a dynamically allocated irq +* @irq: irq number to initialize +*/ +void dynamic_irq_cleanup (unsigned int irq) +{ +dynamic_irq_init_x (irq, false);should be
dynamic_irq_cleanup_xhere.
Quoted 13 lines
kernel/irq/chip.c void dynamic_irq_cleanup(unsigned int ir · line 119
Quoted 11 lines
kernel/irq/chip.c +11
-0+} + +/** +* dynamic_irq_cleanup_keep_chip_data - cleanup a dynamically allocated irq +* @irq: irq number to initialize +* +* does not set irq_to_desc(irq)->chip_data to NULL +*/ +void dynamic_irq_cleanup_keep_chip_data (unsigned int irq) +{ +dynamic_irq_init_x (irq, true);should be
dynamic_irq_cleanup_x
Oops, right. I will fix this up along with the externs as hpa suggested and send again.
What are your thoughts on locking? Does it look OK as is?
Thanks,
Brandon
Re: [PATCH] x86: keep chip_data in create_irq_nr and destroy_irq
Quoted 40 lines · Brandon Philips
Quoted 25 lines · Yinghai Lu
Quoted 11 lines · Brandon Philips
kernel/irq/chip.c +2
-1@@ -77,7 +90,8 @@ void dynamic_irq_cleanup(unsigned int ir } desc->msi_desc =NULL ; desc->handler_data =NULL ; - desc->chip_data =NULL ; +if (!keep_chip_data) + desc->chip_data =NULL ; desc->handle_irq = handle_bad_irq; desc->chip = &no_irq_chip; desc->name =NULL ;kernel/irq/chip.c void dynamic_irq_cleanup(unsigned int ir · line 108
Quoted 10 lines
kernel/irq/chip.c +7
-0@@ -85,6 +99,26 @@ void dynamic_irq_cleanup(unsigned int irraw_spin_unlock_irqrestore (&desc->lock, flags); } +/** +* dynamic_irq_cleanup - cleanup a dynamically allocated irq +* @irq: irq number to initialize +*/ +void dynamic_irq_cleanup (unsigned int irq) +{ +dynamic_irq_init_x (irq, false);should be
dynamic_irq_cleanup_xhere.Quoted 13 lines
kernel/irq/chip.c void dynamic_irq_cleanup(unsigned int ir · line 119
Quoted 11 lines
kernel/irq/chip.c +11
-0+} + +/** +* dynamic_irq_cleanup_keep_chip_data - cleanup a dynamically allocated irq +* @irq: irq number to initialize +* +* does not set irq_to_desc(irq)->chip_data to NULL +*/ +void dynamic_irq_cleanup_keep_chip_data (unsigned int irq) +{ +dynamic_irq_init_x (irq, true);should be
dynamic_irq_cleanup_x
Quoted 6 lines
Oops, right. I will fix this up along with the externs as hpa suggested and send again.
What are your thoughts on locking? Does it look OK as is?
ok to me.
YH
PATCH v2 1/1 x86: keep chip_data in create_irq_nr and destroy_irq
When two drivers are setting up MSI-X at the same time via pci_enable_msix() there is a race. See this dmesg excerpt:
[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereference
at 0000000000000088
As you can see igb and ixgbe are both alternating on create_irq_nr() via pci_enable_msix() in their probe function.
ixgbe: While looping through irq_desc_ptrs[] via create_irq_nr() ixgbe choses irq_desc_ptrs[102] and exits the loop, drops vector_lock and calls dynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL via dynamic_irq_init().
igb: Grabs the vector_lock now and starts looping over irq_desc_ptrs[] via create_irq_nr(). It gets to irq_desc_ptrs[102] and does this:
cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
This hits the NULL deref.
Another possible race exists via pci_disable_msix() in a driver or in the number of error paths that call free_msi_irqs():
destroy_irq()dynamic_irq_cleanup() which sets desc->chip_data = NULL ...race window...desc->chip_data = cfg;
Remove the save and restore code for cfg in create_irq_nr() and destroy_irq() to avoid the NULL.
--- arch/x86/kernel/apic/io_apic.c | 14 +++-------- include/linux/irq.h | 2 + kernel/irq/chip.c | 52 +++++++++++++++++++++++++++++++++-------- 3 files changed, 49 insertions(+), 19 deletions(-)
arch/x86/kernel/apic/io_apic.c +4 -10
@@ -3228,12 +3228,9 @@ unsigned int create_irq_nr(unsigned int }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
@@ -3260,10 +3257,7 @@ void destroy_irq(unsigned int irq)/* store it, in case dynamic_irq_cleanup clear it */ desc =irq_to_desc (irq); - cfg = desc->chip_data; -dynamic_irq_cleanup (irq); -/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags);
include/linux/irq.h +2 -0
@@ -400,7 +400,9 @@ static inline int irq_has_action(unsigne/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq); +void dynamic_irq_cleanup_keep_chip_data (unsigned int irq);/* Set/get chip/data for an IRQ: */ extern int set_irq_chip (unsigned int irq,struct irq_chip *chip);
kernel/irq/chip.c +43 -9
84 lines not on this page · raw message
Re: [PATCH v3] x86: keep chip_data in create_irq_nr and destroy_irq
Version 3: Forgot to refresh patch so destroy_irq() had uninitialized cfg as param to __clear_irq_vector(). Fixed now.
When two drivers are setting up MSI-X at the same time via pci_enable_msix() there is a race. See this dmesg excerpt:
[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereference
at 0000000000000088
As you can see igb and ixgbe are both alternating on create_irq_nr() via pci_enable_msix() in their probe function.
ixgbe: While looping through irq_desc_ptrs[] via create_irq_nr() ixgbe choses irq_desc_ptrs[102] and exits the loop, drops vector_lock and calls dynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL via dynamic_irq_init().
igb: Grabs the vector_lock now and starts looping over irq_desc_ptrs[] via create_irq_nr(). It gets to irq_desc_ptrs[102] and does this:
cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
This hits the NULL deref.
Another possible race exists via pci_disable_msix() in a driver or in the number of error paths that call free_msi_irqs():
destroy_irq()dynamic_irq_cleanup() which sets desc->chip_data = NULL ...race window...desc->chip_data = cfg;
Remove the save and restore code for cfg in create_irq_nr() and destroy_irq() and take the desc->lock when checking the irq_cfg.
--- arch/x86/kernel/apic/io_apic.c | 17 +++---------- include/linux/irq.h | 2 + kernel/irq/chip.c | 52 +++++++++++++++++++++++++++++++++-------- 3 files changed, 50 insertions(+), 21 deletions(-)
arch/x86/kernel/apic/io_apic.c +5 -12
@@ -3228,12 +3228,9 @@ unsigned int create_irq_nr(unsigned int }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
@@ -3255,19 +3252,15 @@ int create_irq(void)void destroy_irq (unsigned int irq) {unsigned long flags; -struct irq_cfg *cfg;struct irq_desc *desc;/* store it, in case dynamic_irq_cleanup clear it */ desc =irq_to_desc (irq); - cfg = desc->chip_data; -dynamic_irq_cleanup (irq); -/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags); -__clear_irq_vector (irq, cfg); +__clear_irq_vector (irq, desc->chip_data);spin_unlock_irqrestore (&vector_lock, flags); }
include/linux/irq.h +2 -0
@@ -400,7 +400,9 @@ static inline int irq_has_action(unsigne/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq); +void dynamic_irq_cleanup_keep_chip_data (unsigned int irq);/* Set/get chip/data for an IRQ: */ extern int set_irq_chip (unsigned int irq,struct irq_chip *chip);
kernel/irq/chip.c +43 -9
84 lines not on this page · raw message
Re: [PATCH v3] x86: keep chip_data in create_irq_nr and destroy_irq
Quoted 14 lines · Brandon Philips
Version 3: Forgot to refresh patch so
destroy_irq()had uninitialized cfg as param to__clear_irq_vector(). Fixed now.arch/x86/kernel/apic/io_apic.c +0
-3@@ -3255,19 +3252,15 @@ int create_irq(void)void destroy_irq (unsigned int irq) {unsigned long flags; -struct irq_cfg *cfg;struct irq_desc *desc;/* store it, in case dynamic_irq_cleanup clear it */ desc =irq_to_desc (irq); - cfg = desc->chip_data; -dynamic_irq_cleanup (irq);
arch/x86/kernel/apic/io_apic.c int create_irq(void) · line 3265
Quoted 10 lines
arch/x86/kernel/apic/io_apic.c +2
-3-/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags); -__clear_irq_vector (irq, cfg); +__clear_irq_vector (irq, desc->chip_data);spin_unlock_irqrestore (&vector_lock, flags); }
==>
@@ -3308,17 +3305,12 @@ void destroy_irq(unsigned int irq)
{
unsigned long flags;
struct irq_cfg *cfg;
- struct irq_desc *desc;
- /* store it, in case dynamic_irq_cleanup clear it */
- desc = irq_to_desc(irq);
- cfg = desc->chip_data;
- dynamic_irq_cleanup(irq);
- /* connect back irq_cfg */
- desc->chip_data = cfg;
+ dynamic_irq_cleanup_keep_chip_data(irq);
free_irte(irq); spin_lock_irqsave(&vector_lock, flags);
+ cfg = irq_to_desc(irq)->chip_data;
__clear_irq_vector(irq, cfg); spin_unlock_irqrestore(&vector_lock, flags);
}
Yinghai
Re: [PATCH v3] x86: keep chip_data in create_irq_nr and destroy_irq
Quoted 9 lines · Yinghai Lu
On 02/05/2010 10:42 PM, Brandon Philips wrote:
arch/x86/kernel/apic/io_apic.c +0
-3@@ -3308,17 +3305,12 @@ void destroy_irq(unsigned int irq) {unsigned long flags;struct irq_cfg *cfg; -struct irq_desc *desc; -/* store it, in case dynamic_irq_cleanup clear it */ - desc =irq_to_desc (irq);
arch/x86/kernel/apic/io_apic.c void destroy_irq(unsigned int irq) · line 3314
Quoted 10 lines
arch/x86/kernel/apic/io_apic.c +2
-4- cfg = desc->chip_data; -dynamic_irq_cleanup (irq); -/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags); + cfg =irq_to_desc (irq)->chip_data;__clear_irq_vector (irq, cfg);
Or even two lines better!
void destroy_irq(unsigned int irq)
{
unsigned long flags;- struct irq_cfg *cfg;
- struct irq_desc *desc;
- /* store it, in case dynamic_irq_cleanup clear it */ - desc = irq_to_desc(irq); - cfg = desc->chip_data; - dynamic_irq_cleanup(irq); - /* connect back irq_cfg */ - desc->chip_data = cfg; + dynamic_irq_cleanup_keep_chip_data(irq);
free_irte(irq);
spin_lock_irqsave(&vector_lock, flags);
- __clear_irq_vector(irq, cfg);
+ __clear_irq_vector(irq, get_irq_chip_data(irq));
spin_unlock_irqrestore(&vector_lock, flags);}
Sigh... I guess I will send the patch again :D
Cheers,
Brandon
Re: [PATCH v4] x86: keep chip_data in create_irq_nr and destroy_irq
In reply to Yinghai LuVersion 4: use get_irq_chip_data() in destroy_irq() to get rid of some local vars.
When two drivers are setting up MSI-X at the same time via pci_enable_msix() there is a race. See this dmesg excerpt:
[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereference
at 0000000000000088
As you can see igb and ixgbe are both alternating on create_irq_nr() via pci_enable_msix() in their probe function.
ixgbe: While looping through irq_desc_ptrs[] via create_irq_nr() ixgbe choses irq_desc_ptrs[102] and exits the loop, drops vector_lock and calls dynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL via dynamic_irq_init().
igb: Grabs the vector_lock now and starts looping over irq_desc_ptrs[] via create_irq_nr(). It gets to irq_desc_ptrs[102] and does this:
cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
This hits the NULL deref.
Another possible race exists via pci_disable_msix() in a driver or in the number of error paths that call free_msi_irqs():
destroy_irq()dynamic_irq_cleanup() which sets desc->chip_data = NULL ...race window...desc->chip_data = cfg;
Remove the save and restore code for cfg in create_irq_nr() and destroy_irq() and take the desc->lock when checking the irq_cfg.
--- arch/x86/kernel/apic/io_apic.c | 20 +++------------ include/linux/irq.h | 2 + kernel/irq/chip.c | 52 +++++++++++++++++++++++++++++++++-------- 3 files changed, 50 insertions(+), 24 deletions(-)
arch/x86/kernel/apic/io_apic.c +5 -15
@@ -3228,12 +3228,9 @@ unsigned int create_irq_nr(unsigned int }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
@@ -3255,19 +3252,12 @@ int create_irq(void)void destroy_irq (unsigned int irq) {unsigned long flags; -struct irq_cfg *cfg; -struct irq_desc *desc; -/* store it, in case dynamic_irq_cleanup clear it */ - desc =irq_to_desc (irq); - cfg = desc->chip_data; -dynamic_irq_cleanup (irq); -/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags); -__clear_irq_vector (irq, cfg); +__clear_irq_vector (irq,get_irq_chip_data (irq));spin_unlock_irqrestore (&vector_lock, flags); }
include/linux/irq.h +2 -0
@@ -400,7 +400,9 @@ static inline int irq_has_action(unsigne/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq); +void dynamic_irq_cleanup_keep_chip_data (unsigned int irq);/* Set/get chip/data for an IRQ: */ extern int set_irq_chip (unsigned int irq,struct irq_chip *chip);
kernel/irq/chip.c +43 -9
84 lines not on this page · raw message
[tip:x86/urgent] x86, irq: Keep chip_data in create_irq_nr and destroy_irq
Commit-ID: eb5b3794062824ba12d883901eea49ea89d0a678 Gitweb: http://git.kernel.org/tip/eb5b3794062824ba12d883901eea49ea89d0a678 Author: Brandon Philips <[email protected]> AuthorDate: Sun, 7 Feb 2010 13:02:50 -0800 Committer: H. Peter Anvin <[email protected]> CommitDate: Thu, 18 Feb 2010 21:53:15 -0800
x86, irq: Keep chip_data in create_irq_nr and destroy_irq
Version 4: use get_irq_chip_data() in destroy_irq() to get rid of some local vars.
When two drivers are setting up MSI-X at the same time via pci_enable_msix() there is a race. See this dmesg excerpt:
[ 85.170610] ixgbe 0000:02:00.1: irq 97 for MSI/MSI-X [ 85.170611] alloc irq_desc for 99 on node -1 [ 85.170613] igb 0000:08:00.1: irq 98 for MSI/MSI-X [ 85.170614] alloc kstat_irqs on node -1 [ 85.170616] alloc irq_2_iommu on node -1 [ 85.170617] alloc irq_desc for 100 on node -1 [ 85.170619] alloc kstat_irqs on node -1 [ 85.170621] alloc irq_2_iommu on node -1 [ 85.170625] ixgbe 0000:02:00.1: irq 99 for MSI/MSI-X [ 85.170626] alloc irq_desc for 101 on node -1 [ 85.170628] igb 0000:08:00.1: irq 100 for MSI/MSI-X [ 85.170630] alloc kstat_irqs on node -1 [ 85.170631] alloc irq_2_iommu on node -1 [ 85.170635] alloc irq_desc for 102 on node -1 [ 85.170636] alloc kstat_irqs on node -1 [ 85.170639] alloc irq_2_iommu on node -1 [ 85.170646] BUG: unable to handle kernel NULL pointer dereference
at 0000000000000088
As you can see igb and ixgbe are both alternating on create_irq_nr() via pci_enable_msix() in their probe function.
ixgbe: While looping through irq_desc_ptrs[] via create_irq_nr() ixgbe choses irq_desc_ptrs[102] and exits the loop, drops vector_lock and calls dynamic_irq_init. Then it sets irq_desc_ptrs[102]->chip_data = NULL via dynamic_irq_init().
igb: Grabs the vector_lock now and starts looping over irq_desc_ptrs[] via create_irq_nr(). It gets to irq_desc_ptrs[102] and does this:
cfg_new = irq_desc_ptrs[102]->chip_data; if (cfg_new->vector != 0) continue;
This hits the NULL deref.
Another possible race exists via pci_disable_msix() in a driver or in the number of error paths that call free_msi_irqs():
destroy_irq()dynamic_irq_cleanup() which sets desc->chip_data = NULL ...race window...desc->chip_data = cfg;
Remove the save and restore code for cfg in create_irq_nr() and destroy_irq() and take the desc->lock when checking the irq_cfg.
LKML-Reference: <[email protected]>
--- arch/x86/kernel/apic/io_apic.c | 20 ++++----------- include/linux/irq.h | 2 + kernel/irq/chip.c | 52 +++++++++++++++++++++++++++++++++------- 3 files changed, 50 insertions(+), 24 deletions(-)
arch/x86/kernel/apic/io_apic.c +5 -15
@@ -3278,12 +3278,9 @@ unsigned int create_irq_nr(unsigned int irq_want, int node) }spin_unlock_irqrestore (&vector_lock, flags); -if (irq >0 ) { -dynamic_irq_init (irq); -/* restore it, in case dynamic_irq_init clear it */ -if (desc_new) - desc_new->chip_data = cfg_new; - } +if (irq >0 ) +dynamic_irq_init_keep_chip_data (irq); +return irq; }
@@ -3305,19 +3302,12 @@ int create_irq(void)void destroy_irq (unsigned int irq) {unsigned long flags; -struct irq_cfg *cfg; -struct irq_desc *desc; -/* store it, in case dynamic_irq_cleanup clear it */ - desc =irq_to_desc (irq); - cfg = desc->chip_data; -dynamic_irq_cleanup (irq); -/* connect back irq_cfg */ - desc->chip_data = cfg; +dynamic_irq_cleanup_keep_chip_data (irq);free_irte (irq);spin_lock_irqsave (&vector_lock, flags); -__clear_irq_vector (irq, cfg); +__clear_irq_vector (irq,get_irq_chip_data (irq));spin_unlock_irqrestore (&vector_lock, flags); }
include/linux/irq.h +2 -0
@@ -400,7 +400,9 @@ static inline int irq_has_action(unsigned int irq)/* Dynamic irq helper functions */ extern void dynamic_irq_init (unsigned int irq); +void dynamic_irq_init_keep_chip_data (unsigned int irq);extern void dynamic_irq_cleanup (unsigned int irq); +void dynamic_irq_cleanup_keep_chip_data (unsigned int irq);/* Set/get chip/data for an IRQ: */ extern int set_irq_chip (unsigned int irq,struct irq_chip *chip);
kernel/irq/chip.c +43 -9
84 lines not on this page · raw message
[tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
In reply to Brandon PhilipsCommit-ID: f6a929eb62ef43cae16ba6d4e2c64b4e430fe7a4 Gitweb: http://git.kernel.org/tip/f6a929eb62ef43cae16ba6d4e2c64b4e430fe7a4 Author: Ingo Molnar <[email protected]> AuthorDate: Fri, 26 Feb 2010 11:17:16 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Fri, 26 Feb 2010 11:17:16 +0100
x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Merge commit aef55d4922 mis-merged io_apic.c so we lost the arch_probe_nr_irqs() method.
This caused subtle boot breakages (udev confusion likely due to missing drivers) with certain configs.
Cc: H. Peter Anvin <[email protected]>
Cc: Yinghai Lu <[email protected]>
LKML-Reference: <[email protected]>
--- arch/x86/kernel/apic/io_apic.c | 22 ++++++++++++++++++++++ 1 files changed, 22 insertions(+), 0 deletions(-)
arch/x86/kernel/apic/io_apic.c +22 -0
@@ -3874,6 +3874,28 @@ void __init probe_nr_irqs_gsi(void)printk (KERN_DEBUG "nr_irqs_gsi: %d\n" , nr_irqs_gsi); } +#ifdef CONFIG_SPARSE_IRQ +int __init arch_probe_nr_irqs (void ) +{ +int nr; + +if (nr_irqs > (NR_VECTORS * nr_cpu_ids)) + nr_irqs =NR_VECTORS * nr_cpu_ids; + + nr = nr_irqs_gsi +8 * nr_cpu_ids; +#if defined(CONFIG_PCI_MSI ) || defined(CONFIG_HT_IRQ ) +/* +* for MSI and HT dyn irq +*/ + nr += nr_irqs_gsi *16 ; +#endif +if (nr < nr_irqs) + nr_irqs = nr; + +return 0 ; +} +#endif +static int __io_apic_set_pci_routing (struct device *dev,int irq,struct io_apic_irq_attr *irq_attr) {
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 46 lines · tip-bot for Ingo Molnar
Commit-ID: f6a929eb62ef43cae16ba6d4e2c64b4e430fe7a4 Gitweb: http://git.kernel.org/tip/f6a929eb62ef43cae16ba6d4e2c64b4e430fe7a4 Author: Ingo Molnar <[email protected]> AuthorDate: Fri, 26 Feb 2010 11:17:16 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Fri, 26 Feb 2010 11:17:16 +0100x86: apic: Fix mismerge, add
arch_probe_nr_irqs()againMerge commit
aef55d4922mis-mergedio_apic.cso we lost thearch_probe_nr_irqs()method.This caused subtle boot breakages (udev confusion likely due to missing drivers) with certain configs.
Cc: H. Peter Anvin <[email protected]>
Cc: Yinghai Lu <[email protected]>
LKML-Reference: <[email protected]>Signed-off-by Ingo Molnar <[email protected]>--- arch/x86/kernel/apic/io_apic.c | 22 ++++++++++++++++++++++ 1 files changed, 22 insertions(+), 0 deletions(-)arch/x86/kernel/apic/io_apic.c +15
-0@@ -3874,6 +3874,28 @@ void __init probe_nr_irqs_gsi(void)printk (KERN_DEBUG "nr_irqs_gsi: %d\n" , nr_irqs_gsi); } +#ifdef CONFIG_SPARSE_IRQ +int __init arch_probe_nr_irqs (void ) +{ +int nr; + +if (nr_irqs > (NR_VECTORS * nr_cpu_ids)) + nr_irqs =NR_VECTORS * nr_cpu_ids; + + nr = nr_irqs_gsi +8 * nr_cpu_ids; +#if defined(CONFIG_PCI_MSI ) || defined(CONFIG_HT_IRQ ) +/* +* for MSI and HT dyn irq +*/ + nr += nr_irqs_gsi *16 ; +#endif
arch/x86/kernel/apic/io_apic.c void __init probe_nr_irqs_gsi(void) · line 3901
Quoted 10 lines
arch/x86/kernel/apic/io_apic.c +7
-0+if (nr < nr_irqs) + nr_irqs = nr; + + return 0; +} +#endif +static int __io_apic_set_pci_routing (struct device *dev,int irq,struct io_apic_irq_attr *irq_attr) {
for x86, with radix tree based irq_to_desc(),
removing arch_probe_nr_irqs is intentional. so we get more irq that could be used.
wonder if the udev for some of your test system have irq number limitation?
YH
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 61 lines · Yinghai Lu
Quoted 46 lines · tip-bot for Ingo Molnar
Commit-ID: f6a929eb62ef43cae16ba6d4e2c64b4e430fe7a4 Gitweb: http://git.kernel.org/tip/f6a929eb62ef43cae16ba6d4e2c64b4e430fe7a4 Author: Ingo Molnar <[email protected]> AuthorDate: Fri, 26 Feb 2010 11:17:16 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Fri, 26 Feb 2010 11:17:16 +0100x86: apic: Fix mismerge, add
arch_probe_nr_irqs()againMerge commit
aef55d4922mis-mergedio_apic.cso we lost thearch_probe_nr_irqs()method.This caused subtle boot breakages (udev confusion likely due to missing drivers) with certain configs.
Cc: H. Peter Anvin <[email protected]>
Cc: Yinghai Lu <[email protected]>
LKML-Reference: <[email protected]>Signed-off-by Ingo Molnar <[email protected]>--- arch/x86/kernel/apic/io_apic.c | 22 ++++++++++++++++++++++ 1 files changed, 22 insertions(+), 0 deletions(-)arch/x86/kernel/apic/io_apic.c +15
-0@@ -3874,6 +3874,28 @@ void __init probe_nr_irqs_gsi(void)printk (KERN_DEBUG "nr_irqs_gsi: %d\n" , nr_irqs_gsi); } +#ifdef CONFIG_SPARSE_IRQ +int __init arch_probe_nr_irqs (void ) +{ +int nr; + +if (nr_irqs > (NR_VECTORS * nr_cpu_ids)) + nr_irqs =NR_VECTORS * nr_cpu_ids; + + nr = nr_irqs_gsi +8 * nr_cpu_ids; +#if defined(CONFIG_PCI_MSI ) || defined(CONFIG_HT_IRQ ) +/* +* for MSI and HT dyn irq +*/ + nr += nr_irqs_gsi *16 ; +#endif arch/x86/kernel/apic/io_apic.c void __init probe_nr_irqs_gsi(void) · line 3901
Quoted 10 lines
arch/x86/kernel/apic/io_apic.c +7
-0+if (nr < nr_irqs) + nr_irqs = nr; + + return 0; +} +#endif +static int __io_apic_set_pci_routing (struct device *dev,int irq,struct io_apic_irq_attr *irq_attr) {for x86, with radix tree based
irq_to_desc(),
removingarch_probe_nr_irqsis intentional. so we get more irq that could be used.
Quoted 1 line
wonder if the udev for some of your test system have irq number limitation?
was ancient udev: udev-095-17.fc6.
Ingo
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 9 lines · Ingo Molnar
Quoted 5 lines · Yinghai Lu
for x86, with radix tree based
irq_to_desc(),
removingarch_probe_nr_irqsis intentional. so we get more irq that could be used.wonder if the udev for some of your test system have irq number limitation?
was ancient udev: udev-095-17.fc6.
Something doesn't add up. Nowhere in the udev source is there a single mention of irq.
gsi have fixed interrupt numbers so that would not change.
The dynamic irqs are allocated starting from the high gsi and working up.
The irq numbers that get allocated should not have changed, unless this was actually a bug fix in this configuration.
The other possibility is that somehow arch_probe_nr_irqs() was returning a number greater than NR_IRQS.
Ingo do you have any idea what NR_IRQS or nr_irqs were/are on that failing machine?
Eric
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 20 lines · Eric W. Biederman
Quoted 9 lines · Ingo Molnar
Quoted 5 lines · Yinghai Lu
for x86, with radix tree based
irq_to_desc(),
removingarch_probe_nr_irqsis intentional. so we get more irq that could be used.wonder if the udev for some of your test system have irq number limitation?
was ancient udev: udev-095-17.fc6.
Something doesn't add up. Nowhere in the udev source is there a single mention of irq.
gsi have fixed interrupt numbers so that would not change.
The dynamic irqs are allocated starting from the high gsi and working up.
Quoted 8 lines
The irq numbers that get allocated should not have changed, unless this was actually a bug fix in this configuration.
The other possibility is that somehow
arch_probe_nr_irqs()was returning a number greater thanNR_IRQS.Ingo do you have any idea what
NR_IRQSornr_irqswere/are on that failing machine?
Sorry, not - and the merge window doesnt leave much time to revisit the problem right now.
But the failures were very real and 100% caused by this: they resulted in non-existent /dev/sda* nodes and resulting fsck failure by rc.
Thanks,
Ingo
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 8 lines · Ingo Molnar
Quoted 2 lines
Ingo do you have any idea what
NR_IRQSornr_irqswere/are on that failing machine?Sorry, not - and the merge window doesnt leave much time to revisit the problem right now.
But the failures were very real and 100% caused by this: they resulted in non-existent
/dev/sda*nodes and resulting fsck failure by rc.
I have looked it over a second time and I have convinced myself that arch_probe_nr_irqs will in the worst case reduce nr_irqs, and never increase it beyond NR_IRQS. So this revert (keeping arch_probe_nr_irqs) is safe.
It makes little sense that a larger nr_irqs would be a problem, but clearly there are assumptions somewhere that we still need to remove.
Eric
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
In reply to Yinghai LuQuoted 5 lines · Yinghai Lu
for x86, with radix tree based
irq_to_desc(),
removingarch_probe_nr_irqsis intentional. so we get more irq that could be used.
This sounds interesting for the Xen dom0 use case where we can have essentially arbitrary numbers of interrupt sources.
Is there a tree somewhere that I can be looking at?
Ian
Signature · 4 lines
Ian Campbell Current Noise: Rammstein - Morgenstern Money will say more in one moment than the most eloquent lover can in years.
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 11 lines · Ian Campbell
Quoted 5 lines · Yinghai Lu
for x86, with radix tree based
irq_to_desc(),
removingarch_probe_nr_irqsis intentional. so we get more irq that could be used.This sounds interesting for the Xen dom0 use case where we can have essentially arbitrary numbers of interrupt sources.
Is there a tree somewhere that I can be looking at?
Ingo's x86 tip tree I imagine. I think this patch is slated for coming in sometime this merge window so you should be able to see it in 2.6.34-rc1.
This is a truly frustrating question. At this point all of the practical limitations are in the Xen code itself, and it has been that way for the last couple of years. Looking at the x86_64 tree it looks like the core work went in at the end of 2006 about 2.6.19. Last I looked the Xen dom0 kernel was stuck at the kernel just before the irq scaling changes went in. Every time I look the Xen code is doing something stupid and unnecessary that makes scaling to large numbers of irqs hard.
The changes YH is working on are the last couple of changes so that we can seriously remove NR_IRQs and not have to pay a penalty of a large static array on small machines when we have a kernel built to support large machines and a large number of interrupts.
As of 2.6.33 the limitations in DomU support are:
- xen_evtchn_do_upcall starts with the irq number instead of the irq_desc, and happens to unnecessarily call into arch specific code.
- Xen has an array irq_info[NR_IRQS] one of the last static arrays sized at NR_IRQs in the entire kernel.
One of the big reasons Xen dom0 irq support was not merged was because merging it would effectively be a revert of the irq scaling work.
If you can fix the Xen code so it isn't dragging the rest of the kernel down when it comes to large numbers of irqs more power to you. For now I will continue to write little patches every once in a while that bonk Xen domU on the head when something Xen domU is doing becomes a bottleneck for the rest of the kernel.
Eric
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 5 lines · Eric W. Biederman
As of 2.6.33 the limitations in DomU support are:
-xen_evtchn_do_upcallstarts with the irq number instead of theirq_desc, and happens to unnecessarily call into arch specific code.
I saw a patch to fix this one recently, "xen: Remove unnecessary arch specific xen irq functions.", right?
Quoted 2 lines
- Xen has an array irq_info[NR_IRQS] one of the last static arrays sized at
NR_IRQsin the entire kernel.
Hopefully the same info as is in that array could (and indeed should) be instead stored in irq_desc->chip_data. Would you object to arch_init_copy_chip_data and arch_free_chip_data becoming function pointers within the struct irq_chip?
Quoted 2 lines
If you can fix the Xen code so it isn't dragging the rest of the kernel down when it comes to large numbers of irqs more power to you.
If you know of other aspects of the Xen code where this is the case (or find them in the future) please let me know, I'll do my best to fix them.
Ian.
Signature · 4 lines
Ian Campbell To the systems programmer, users and applications serve only to provide a test load.
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 9 lines · Ian Campbell
Quoted 5 lines · Eric W. Biederman
As of 2.6.33 the limitations in DomU support are:
-xen_evtchn_do_upcallstarts with the irq number instead of theirq_desc, and happens to unnecessarily call into arch specific code.I saw a patch to fix this one recently, "xen: Remove unnecessary arch specific xen irq functions.", right?
Yes.
You probably want to modify evtnchn_to_irq to return an irq_desc.
It is going to take a bit but our next big step for the irq methods is to make them all take struct irq_desc pointers instead of unsigned int irq, so we don't have to repeat the lookups.
Quoted 7 lines
Quoted 2 lines
- Xen has an array irq_info[NR_IRQS] one of the last static arrays sized at
NR_IRQsin the entire kernel.Hopefully the same info as is in that array could (and indeed should) be instead stored in
irq_desc->chip_data. Would you object toarch_init_copy_chip_dataandarch_free_chip_databecoming function pointers within the structirq_chip?
No objections. Now that I see those methods it looks like they always should have been in irq_chip.
Quoted 6 lines
Quoted 2 lines
If you can fix the Xen code so it isn't dragging the rest of the kernel down when it comes to large numbers of irqs more power to you.
If you know of other aspects of the Xen code where this is the case (or find them in the future) please let me know, I'll do my best to fix them.
Good to hear.
Right now I am being a bit of a jack in the box reviewer. I don't have time to work on the irq code right now but I occasionally pop out of my box and review the code and try to keep it from descending into an unmaintainable disaster.
Eric
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Quoted 5 lines · Eric W. Biederman
Quoted 1 line · Ian Campbell
On Mon, 2010-03-01 at 10:34 -0800, Eric W. Biederman wrote:
It is going to take a bit but our next big step for the irq methods is to make them all take struct
irq_descpointers instead of unsigned int irq, so we don't have to repeat the lookups.
That should hit 2.6.35.
Quoted 10 lines
Quoted 7 lines
Quoted 2 lines
- Xen has an array irq_info[NR_IRQS] one of the last static arrays sized at
NR_IRQsin the entire kernel.Hopefully the same info as is in that array could (and indeed should) be instead stored in
irq_desc->chip_data. Would you object toarch_init_copy_chip_dataandarch_free_chip_databecoming function pointers within the structirq_chip?No objections. Now that I see those methods it looks like they always should have been in
irq_chip.
Agreed.
Thanks,
tglx
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
In reply to Eric W. BiedermanQuoted 4 lines · Eric W. Biederman
Quoted 4 lines
Quoted 3 lines
- Xen has an array irq_info[NR_IRQS] one of the last static arrays sized at
NR_IRQsin the entire kernel.
Quoted 8 lines
Quoted 1 line
Hopefully the same info as is in that array could (and indeed
should) be
Quoted 3 lines
instead stored in
irq_desc->chip_data. Would you object toarch_init_copy_chip_dataandarch_free_chip_databecoming function pointers within the structirq_chip?No objections. Now that I see those methods it looks like they always should have been in
irq_chip.
The following changes since commit c0a2d57e753717cc893b4b38bfb351c7f19469c3:
Ingo Molnar (1):
Merge branch 'perf/urgent'
are available in the git repository at:
git://xenbits.xensource.com/people/ianc/linux-2.6.git for-x86/irq
Ian Campbell (3):
x86: namespace some I/O APIC related structures and functions.
irq: move some interrupt arch_* functions into struct irq_chip.
x86: irq_desc->chip_data is always correct whether or not SPARSE_IRQ is enabled.
arch/powerpc/kernel/irq.c | 2 +- arch/x86/include/asm/hw_irq.h | 25 +++-- arch/x86/kernel/apic/io_apic.c | 233 ++++++++++++++++++++++++--------------- arch/x86/kernel/uv_irq.c | 19 ++-- include/linux/interrupt.h | 2 +- include/linux/irq.h | 12 ++- kernel/irq/handle.c | 2 +- kernel/irq/numa_migrate.c | 12 ++- kernel/softirq.c | 3 +- 9 files changed, 195 insertions(+), 115 deletions(-)
Signature · 6 lines
Ian Campbell Current Noise: Mistress - Lord Worm Nothing is easier than to denounce the evildoer; nothing is more difficult than to understand him. -- Fyodor Dostoevski
PATCH 1/1 irq: move some interrupt arch_* functions into struct irq_chip.
From: Ian Campbell <[email protected]>
Move arch_init_copy_chip_data and arch_free_chip_data into function pointers in struct irq_chip since they operate on irq_desc->chip_data.
arch_init_chip_data cannot be moved into struct irq_chip at this time because irq_desc->chip is not known at the time the irq_desc is setup. For now rename arch_init_chip_data to arch_init_irq_desc (for PowerPC, the only other user, whose usage better matches the new name) and on x86 convert arch_init_chip_data to ioapic_init_chip_data and call this whenever the IO APIC code allocates a new IRQ.
I've retained the chip_data behaviour for uv_irq although it isn't clear to me if these interrupt types support migration or how closely related to the APIC modes they really are. If it weren't for this the ioapic_{init,copy,free}_chip_data functions could be static to io_apic.c.
I've tested by booting on a 64 bit system, but it's not clear to me what actions I need to take to actually exercise some of these code paths.
--- arch/powerpc/kernel/irq.c | 2 +- arch/x86/include/asm/hw_irq.h | 11 +++++++- arch/x86/kernel/apic/io_apic.c | 58 +++++++++++++++++++++++++++++++++++++--- arch/x86/kernel/uv_irq.c | 5 +++ include/linux/interrupt.h | 2 +- include/linux/irq.h | 12 +++++--- kernel/irq/handle.c | 2 +- kernel/irq/numa_migrate.c | 12 +++++++- kernel/softirq.c | 3 +- 9 files changed, 91 insertions(+), 16 deletions(-)
arch/powerpc/kernel/irq.c +1 -1
@@ -1088,7 +1088,7 @@ int arch_early_irq_init(void)return 0 ; } -int arch_init_chip_data (struct irq_desc *desc,int node) +int arch_init_irq_desc (struct irq_desc *desc,int node) { desc->status |=IRQ_NOREQUEST ;return 0 ;
arch/x86/include/asm/hw_irq.h +10 -1
@@ -20,9 +20,9 @@#include <linux/percpu.h> #include <linux/profile.h> #include <linux/smp.h> +#include <linux/irq.h> #include <asm/atomic.h> -#include <asm/irq.h> #include <asm/sections.h> /* Interrupt handlers registered during init_IRQ */
@@ -61,6 +61,15 @@ extern void init_VISWS_APIC_irqs(void);extern void setup_IO_APIC (void );extern void disable_IO_APIC (void ); +extern int ioapic_init_chip_data (struct irq_desc *desc,int node); + +#ifdef CONFIG_SPARSE_IRQ +extern void ioapic_copy_chip_data (struct irq_desc *old_desc, +struct irq_desc *desc,int node); +extern void ioapic_free_chip_data (struct irq_desc *old_desc, +struct irq_desc *desc); +#endif +struct io_apic_irq_attr {int ioapic;int ioapic_pin;
arch/x86/kernel/apic/io_apic.c +54 -4
160 lines not on this page · raw message
arch/x86/kernel/uv_irq.c +5 -0
@@ -55,6 +55,11 @@ struct irq_chip uv_irq_chip = { .eoi = uv_ack_apic, .end = uv_noop, .set_affinity = uv_set_irq_affinity, + +#ifdef CONFIG_SPARSE_IRQ + .copy_chip_data = ioapic_copy_chip_data, + .free_chip_data = ioapic_free_chip_data, +#endif };/*
include/linux/interrupt.h +1 -1
@@ -611,6 +611,6 @@ struct irq_desc;extern int early_irq_init (void );extern int arch_probe_nr_irqs (void );extern int arch_early_irq_init (void ); -extern int arch_init_chip_data (struct irq_desc *desc,int node); +extern int arch_init_irq_desc (struct irq_desc *desc,int node);#endif
include/linux/irq.h +8 -4
@@ -131,6 +131,14 @@ struct irq_chip {void (*bus_lock)(unsigned int irq);void (*bus_sync_unlock)(unsigned int irq); +/* for move_irq_desc */ +#ifdef CONFIG_SPARSE_IRQ +void (*copy_chip_data)(struct irq_desc *old_desc, +struct irq_desc *desc,int node); +void (*free_chip_data)(struct irq_desc *old_desc, +struct irq_desc *desc); +#endif +/* Currently used only by UML, might disappear one day.*/ #ifdef CONFIG_IRQ_RELEASE_METHOD void (*release)(unsigned int irq,void *dev_id);
@@ -208,10 +216,6 @@ struct irq_desc {const char *name; }____cacheline_internodealigned_in_smp ; -extern void arch_init_copy_chip_data (struct irq_desc *old_desc, -struct irq_desc *desc,int node); -extern void arch_free_chip_data (struct irq_desc *old_desc,struct irq_desc *desc); -#ifndef CONFIG_SPARSE_IRQ extern struct irq_desc irq_desc[NR_IRQS ];#endif
kernel/irq/handle.c +1 -1
@@ -120,7 +120,6 @@ static void init_one_irq_desc(int irq, struct irq_desc *desc, int node)BUG_ON (1 ); }init_desc_masks (desc); -arch_init_chip_data (desc, node); }/*
@@ -228,6 +227,7 @@ struct irq_desc * __ref irq_to_desc_alloc_node(unsigned int irq, int node)BUG_ON (1 ); }init_one_irq_desc (irq, desc, node); +arch_init_irq_desc (desc, node);set_irq_desc (irq, desc);
kernel/irq/numa_migrate.c +10 -2
@@ -47,7 +47,8 @@ static bool init_copy_one_irq_desc(int irq, struct irq_desc *old_desc,lockdep_set_class (&desc->lock, &irq_desc_lock_class);init_copy_kstat_irqs (old_desc, desc, node, nr_cpu_ids);init_copy_desc_masks (old_desc, desc); -arch_init_copy_chip_data (old_desc, desc, node); +if (desc->chip->copy_chip_data) + desc->chip->copy_chip_data (old_desc, desc, node);return true; }
@@ -55,7 +56,8 @@ static void free_one_irq_desc(struct irq_desc *old_desc, struct irq_desc *desc) {free_kstat_irqs (old_desc, desc);free_desc_masks (old_desc, desc); -arch_free_chip_data (old_desc, desc); +if (desc->chip->free_chip_data) + desc->chip->free_chip_data (old_desc, desc); } static struct irq_desc *__real_move_irq_desc(struct irq_desc *old_desc,
@@ -107,9 +109,15 @@ out_unlock:struct irq_desc *move_irq_desc (struct irq_desc *desc,int node) { +/* those static or target node is -1, do not move them */ if (desc->irq <NR_IRQS_LEGACY || node ==-1 )return desc; +/* IRQ chip does not support movement */ +if (desc->chip_data && + (desc->chip->copy_chip_data ==NULL || + desc->chip->free_chip_data ==NULL )) +return desc;if (desc->node != node) desc =__real_move_irq_desc (desc, node);
kernel/softirq.c +1 -2
@@ -895,8 +895,7 @@ int __init __weak arch_early_irq_init(void) {return 0 ; } - -int __weak arch_init_chip_data (struct irq_desc *desc,int node) +int __weak arch_init_irq_desc (struct irq_desc *desc,int node) {return 0 ; }
Signature · 1 line
1.5.6.5
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 7 lines · On Wed, 2010-03-10 at 10:55 +0000, [email protected] wrote:
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.
One idea I had to improve this was to add a struct irq_chip * as a parameter to irq_to_desc_alloc_node. The new parameter potentially could be NULL for current behaviour. Does that sound like a reasonable approach?
Ian.
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 9 lines · Ian Campbell
Quoted 7 lines · On Wed, 2010-03-10 at 10:55 +0000, [email protected] wrote:
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.
Quoted 4 lines
One idea I had to improve this was to add a struct
irq_chip* as a parameter toirq_to_desc_alloc_node. The new parameter potentially could be NULL for current behaviour. Does that sound like a reasonable approach?
I don't follow why we have the restriction that irq_to_desc_alloc_node must call arch_init_chip_data. Assuming that requirement to call arch_init_chip_data is valid, passing something into init_one_irq_desc seems appropriate.
Eric
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 11 lines · Eric W. Biederman
Quoted 9 lines · Ian Campbell
Quoted 7 lines · On Wed, 2010-03-10 at 10:55 +0000, [email protected] wrote:
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.
Quoted 8 lines
Quoted 4 lines
One idea I had to improve this was to add a struct
irq_chip* as a parameter toirq_to_desc_alloc_node. The new parameter potentially could be NULL for current behaviour. Does that sound like a reasonable approach?I don't follow why we have the restriction that
irq_to_desc_alloc_nodemust callarch_init_chip_data. Assuming that requirement to callarch_init_chip_datais valid, passing something intoinit_one_irq_descseems appropriate.
Yes, I suspect that could also be made to work.
The lifecycle of the irq_desc and chip_data isn't really clear to me -- I guess once allocated an irq_desc never gets freed (at least currently)? The associated chip_data can be freed on migrate and replaced with a new one, but is not freed otherwise.
My concern is that if the caller asks for an IRQ which already exists (is that valid?) then you will get that existing irq_desc back, including its existing chip_data, which potentially leaks the new one which was passed in. Or is it the case that the only way this could happen would be for legacy IRQs? In which case perhaps it is simply invalid to pass a new chip data in for such an IRQ.
Ian.
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 21 lines · Ian Campbell
Quoted 11 lines · Eric W. Biederman
Quoted 9 lines · Ian Campbell
Quoted 7 lines · On Wed, 2010-03-10 at 10:55 +0000, [email protected] wrote:
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.Quoted 8 lines
Quoted 4 lines
One idea I had to improve this was to add a struct
irq_chip* as a parameter toirq_to_desc_alloc_node. The new parameter potentially could be NULL for current behaviour. Does that sound like a reasonable approach?I don't follow why we have the restriction that
irq_to_desc_alloc_nodemust callarch_init_chip_data. Assuming that requirement to callarch_init_chip_datais valid, passing something intoinit_one_irq_descseems appropriate.
Quoted 6 lines
Yes, I suspect that could also be made to work.
The lifecycle of the
irq_descandchip_dataisn't really clear to me -- I guess once allocated anirq_descnever gets freed (at least currently)? The associatedchip_datacan be freed on migrate and replaced with a new one, but is not freed otherwise.
Yes. That actually looks like a bug.
Quoted 6 lines
My concern is that if the caller asks for an IRQ which already exists (is that valid?) then you will get that existing
irq_descback, including its existingchip_data, which potentially leaks the new one which was passed in. Or is it the case that the only way this could happen would be for legacy IRQs? In which case perhaps it is simply invalid to pass a new chip data in for such an IRQ.
The only irqs that should be allocated/freed are probably the msi irqs as those are the only ones that dynamically come and go in the system.
Unfortunately there appears to be a bigger mess here than first appeared.
Eric
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
In reply to [email protected]Quoted 18 lines · On Wed, Mar 10, 2010 at 2:55 AM, <[email protected]> wrote:
From: Ian Campbell <[email protected]>
Move
arch_init_copy_chip_dataandarch_free_chip_datainto function pointers in structirq_chipsince they operate onirq_desc->chip_data.
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.I've retained the
chip_databehaviour foruv_irqalthough it isn't clear to me if these interrupt types support migration or how closely related to the APIC modes they really are. If it weren't for this the ioapic_{init,copy,free}_chip_data functions could be static toio_apic.c.
Quoted 4 lines
I've tested by booting on a 64 bit system, but it's not clear to me what actions I need to take to actually exercise some of these code paths.
can you just add another pointer field in irq_desc?
some kind of *irq_info etc.
YH
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 19 lines · Yinghai Lu
Quoted 18 lines · On Wed, Mar 10, 2010 at 2:55 AM, <[email protected]> wrote:
From: Ian Campbell <[email protected]>
Move
arch_init_copy_chip_dataandarch_free_chip_datainto function pointers in structirq_chipsince they operate onirq_desc->chip_data.
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.I've retained the
chip_databehaviour foruv_irqalthough it isn't clear to me if these interrupt types support migration or how closely related to the APIC modes they really are. If it weren't for this the ioapic_{init,copy,free}_chip_data functions could be static toio_apic.c.
Quoted 8 lines
Quoted 4 lines
I've tested by booting on a 64 bit system, but it's not clear to me what actions I need to take to actually exercise some of these code paths.
can you just add another pointer field in
irq_desc?some kind of
*irq_infoetc.
I think I don't understand what you are suggesting.
There is already a pointer for irq_chip specific use i.e. irq_desc->chip_data. This patchset is just about ensuring that the field really is available to any chip implementation rather than just assuming it is always used for the acpi chip types (on x86 at least).
Does adding a second pointer with the same (intended?) semantics as the existing one buy us anything?
Ian.
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 25 lines · Ian Campbell
Quoted 24 lines · Yinghai Lu
Quoted 22 lines · On Wed, Mar 10, 2010 at 2:55 AM, <[email protected]> wrote:
From: Ian Campbell <[email protected]>
Move
arch_init_copy_chip_dataandarch_free_chip_datainto function pointers in structirq_chipsince they operate onirq_desc->chip_data.
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.I've retained the
chip_databehaviour foruv_irqalthough it isn't clear to me if these interrupt types support migration or how closely related to the APIC modes they really are. If it weren't for this the ioapic_{init,copy,free}_chip_data functions could be static toio_apic.c.I've tested by booting on a 64 bit system, but it's not clear to me what actions I need to take to actually exercise some of these code paths.
Quoted 5 lines
Quoted 3 lines
can you just add another pointer field in
irq_desc?some kind of
*irq_infoetc.I think I don't understand what you are suggesting.
YH another field doesn't make much sense. Xen is a bizarre subarch with an incompatible irq model. Xen simply needs the ability to handle the entire lifetime of an irq_chip.
All we need between the Xen and the rest of x86 is a convention
so that we never manage the same irqs. At least for domU we are
in an either/or situation so I don't see even that being a problem.
Quoted 4 lines
There is already a pointer for
irq_chipspecific use i.e.irq_desc->chip_data. This patchset is just about ensuring that the field really is available to any chip implementation rather than just assuming it is always used for the acpi chip types (on x86 at least).
Ian Xen in this sense is simply not x86. irq_cfg is not acpi or ioapic or anything but x86 specific. It has everything to do with having a per cpu vector table of 256 entries and architecturally receiving a vector number when an interrupt is fired.
It totally makes sense for Xen to do something different because architecturally it has a completely different irq subsystem.
At the same time let's not pretend that the reason for this is anything except that Xen has a completely different notion of interrupt delivery than the rest of x86 and so it is it's own bizarre subarch.
This is not a case where you simply need a driver because something is a bit different but fits into the existing model.
So the best solution here seems to be a parameter that we pass into irq_to_desc_alloc_node that does what is needed. The second best would be to have arch_init_chip_data to call something likeplatfrom_init_chip_data(). But I think we can avoid that in
this case.
Eric
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 9 lines · Eric W. Biederman
Ian Xen in this sense is simply not x86.
irq_cfgis not acpi or ioapic or anything but x86 specific. It has everything to do with having a per cpu vector table of 256 entries and architecturally receiving a vector number when an interrupt is fired.It totally makes sense for Xen to do something different because architecturally it has a completely different irq subsystem.
OK, so that sounds like you would like the same patchset but without the irq_cfg renaming? or potentially with renaming to x86_blah instead (I'll rework to your preference).
Ian.
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 8 lines · Ian Campbell
Quoted 7 lines · Eric W. Biederman
Ian Xen in this sense is simply not x86.
irq_cfgis not acpi or ioapic or anything but x86 specific. It has everything to do with having a per cpu vector table of 256 entries and architecturally receiving a vector number when an interrupt is fired.
Quoted 6 lines
Quoted 2 lines
It totally makes sense for Xen to do something different because architecturally it has a completely different irq subsystem.
OK, so that sounds like you would like the same patchset but without the
irq_cfgrenaming? or potentially with renaming tox86_blahinstead (I'll rework to your preference).
Currently the renaming really makes it unclear what you are doing and for some reason the description of the renaming rubbed me the wrong way.
Eric
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 13 lines · Eric W. Biederman
Quoted 11 lines · Ian Campbell
Quoted 9 lines · Eric W. Biederman
Ian Xen in this sense is simply not x86.
irq_cfgis not acpi or ioapic or anything but x86 specific. It has everything to do with having a per cpu vector table of 256 entries and architecturally receiving a vector number when an interrupt is fired.It totally makes sense for Xen to do something different because architecturally it has a completely different irq subsystem.
Quoted 6 lines
Quoted 3 lines
OK, so that sounds like you would like the same patchset but without the
irq_cfgrenaming? or potentially with renaming tox86_blahinstead (I'll rework to your preference).Currently the renaming really makes it unclear what you are doing and for some reason the description of the renaming rubbed me the wrong way.
Sorry, I started off a bit confused and then totally misunderstood what related to what and I think that came through in the description.
I'll respin without the first patch.
Ian.
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
In reply to Eric W. BiedermanQuoted 4 lines · Eric W. Biederman
All we need between the Xen and the rest of x86 is a convention
so that we never manage the same irqs. At least for domU we are
in an either/or situation so I don't see even that being a problem.
Dom0 too. This is part of the work implementing what we discussed a while back - Xen now completely owns the local and IO apics, so dom0 only deals with Xen, not the hardware. Xen has a completely different interrupt setup path, but at least it isn't a mishmash of Xen stuff and native APIC stuff.
J
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
In reply to Ian CampbellQuoted 31 lines · Ian Campbell
Quoted 19 lines · Yinghai Lu
Quoted 18 lines · On Wed, Mar 10, 2010 at 2:55 AM, <[email protected]> wrote:
From: Ian Campbell <[email protected]>
Move
arch_init_copy_chip_dataandarch_free_chip_datainto function pointers in structirq_chipsince they operate onirq_desc->chip_data.
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.I've retained the
chip_databehaviour foruv_irqalthough it isn't clear to me if these interrupt types support migration or how closely related to the APIC modes they really are. If it weren't for this the ioapic_{init,copy,free}_chip_data functions could be static toio_apic.c.Quoted 8 lines
Quoted 4 lines
I've tested by booting on a 64 bit system, but it's not clear to me what actions I need to take to actually exercise some of these code paths.
can you just add another pointer field in
irq_desc?some kind of
*irq_infoetc.I think I don't understand what you are suggesting.
Quoted 7 lines
There is already a pointer for
irq_chipspecific use i.e.irq_desc->chip_data. This patchset is just about ensuring that the field really is available to any chip implementation rather than just assuming it is always used for the acpi chip types (on x86 at least).Does adding a second pointer with the same (intended?) semantics as the existing one buy us anything?
#ifdef CONFIG_INTR_REMAP
struct irq_2_iommu *irq_2_iommu;
#endif
struct irq_chip *chip;
struct msi_desc *msi_desc;we already have that for irq_2_iommu and msi_desc
YH
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
Quoted 41 lines · Yinghai Lu
Quoted 31 lines · Ian Campbell
Quoted 19 lines · Yinghai Lu
Quoted 18 lines · On Wed, Mar 10, 2010 at 2:55 AM, <[email protected]> wrote:
From: Ian Campbell <[email protected]>
Move
arch_init_copy_chip_dataandarch_free_chip_datainto function pointers in structirq_chipsince they operate onirq_desc->chip_data.
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.I've retained the
chip_databehaviour foruv_irqalthough it isn't clear to me if these interrupt types support migration or how closely related to the APIC modes they really are. If it weren't for this the ioapic_{init,copy,free}_chip_data functions could be static toio_apic.c.Quoted 8 lines
Quoted 4 lines
I've tested by booting on a 64 bit system, but it's not clear to me what actions I need to take to actually exercise some of these code paths.
can you just add another pointer field in
irq_desc?some kind of
*irq_infoetc.I think I don't understand what you are suggesting.
Quoted 7 lines
There is already a pointer for
irq_chipspecific use i.e.irq_desc->chip_data. This patchset is just about ensuring that the field really is available to any chip implementation rather than just assuming it is always used for the acpi chip types (on x86 at least).Does adding a second pointer with the same (intended?) semantics as the existing one buy us anything?
Quoted 7 lines
#ifdef CONFIG_INTR_REMAP struct irq_2_iommu *irq_2_iommu; #endif struct irq_chip *chip; struct msi_desc *msi_desc;we already have that for
irq_2_iommuandmsi_desc
Those are at different levels of the hierarchy. Adding another pointer for Xen is like having a different iommu and so adding another pointer to handle that kind of iommu.
Eric
Re: [PATCH] irq: move some interrupt arch_* functions into struct irq_chip.
In reply to [email protected]Quoted 11 lines · On Wed, 2010-03-10 at 10:55 +0000, [email protected] wrote:
From: Ian Campbell <[email protected]>
Move
arch_init_copy_chip_dataandarch_free_chip_datainto function pointers in structirq_chipsince they operate onirq_desc->chip_data.
arch_init_chip_datacannot be moved into structirq_chipat this time becauseirq_desc->chipis not known at the time theirq_descis setup. For now renamearch_init_chip_datatoarch_init_irq_desc(for PowerPC, the only other user, whose usage better matches the new name) and on x86 convertarch_init_chip_datatoioapic_init_chip_dataand call this whenever the IO APIC code allocates a new IRQ.
Ack on the name change, it should be called arch_init_irq_desc(), the existing name clearly comes from the fact that sparse IRQ was implemented first on x86, and on x86 that routine init's the chip data for a new irq_desc.
But semantically arch_init_irq_desc() is the right name, I was just too lazy to change it when I enabled sparse IRQ for powerpc.
Can't comment on the rest of the patch.
cheers
PATCH 1/1 x86: namespace some I/O APIC related structures and functions.
In reply to Ian CampbellFrom: Ian Campbell <[email protected]>
It is not obvious that certain functions relate specifically to the I/O APIC. Rename structures and functions as follows:
struct irq_cfg -> struct ioapic_irq_cfg irq_cfg() -> ioapic_irq_cfg() assign_irq_vector() -> ioapic_assign_irq_vector() send_cleanup_vector() -> ioapic_send_cleanup_vector()
There is a slight preference towards ioapic vs io_apic in the current code (563 occurances vs 146) so I went with that.
--- arch/x86/include/asm/hw_irq.h | 14 ++-- arch/x86/kernel/apic/io_apic.c | 174 +++++++++++++++++++++------------------- arch/x86/kernel/uv_irq.c | 14 ++-- 3 files changed, 106 insertions(+), 96 deletions(-)
arch/x86/include/asm/hw_irq.h +8 -6
@@ -83,7 +83,7 @@ static inline void set_io_apic_irq_attr(struct io_apic_irq_attr *irq_attr,* * Most irqs are mapped 1:1 with pins. */ -struct irq_cfg { +struct ioapic_irq_cfg {struct irq_pin_list *irq_2_pin;cpumask_var_t domain;cpumask_var_t old_domain;
@@ -91,13 +91,15 @@ struct irq_cfg { u8 move_in_progress : 1; }; -extern struct irq_cfg *irq_cfg (unsigned int ); -extern int assign_irq_vector (int ,struct irq_cfg *,const struct cpumask *); -extern void send_cleanup_vector (struct irq_cfg *); +extern struct ioapic_irq_cfg *ioapic_irq_cfg (unsigned int ); +extern int ioapic_assign_irq_vector (int ,struct ioapic_irq_cfg *, +const struct cpumask *); +extern void ioapic_send_cleanup_vector (struct ioapic_irq_cfg *);struct irq_desc ; -extern unsigned int set_desc_affinity (struct irq_desc *,const struct cpumask *, -unsigned int *dest_id); +extern unsigned int ioapic_set_desc_affinity (struct irq_desc *, +const struct cpumask *, +unsigned int *dest_id);extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin,struct io_apic_irq_attr *irq_attr);extern void setup_ioapic_dest (void );
arch/x86/kernel/apic/io_apic.c +91 -83
614 lines not on this page · raw message
arch/x86/kernel/uv_irq.c +7 -7
@@ -144,7 +144,7 @@ arch_enable_uv_irq(char *irq_name, unsigned int irq, int cpu, int mmr_blade, {const struct cpumask *eligible_cpu =cpumask_of (cpu);struct irq_desc *desc =irq_to_desc (irq); -struct irq_cfg *cfg; +struct ioapic_irq_cfg *cfg;int mmr_pnode;unsigned long mmr_value;struct uv_IO_APIC_route_entry *entry;
@@ -153,9 +153,9 @@ arch_enable_uv_irq(char *irq_name, unsigned int irq, int cpu, int mmr_blade,BUILD_BUG_ON (sizeof (struct uv_IO_APIC_route_entry ) !=sizeof (unsigned long )); - cfg =irq_cfg (irq); + cfg =ioapic_irq_cfg (irq); - err =assign_irq_vector (irq, cfg, eligible_cpu); + err =ioapic_assign_irq_vector (irq, cfg, eligible_cpu);if (err !=0 )return err;
@@ -181,7 +181,7 @@ arch_enable_uv_irq(char *irq_name, unsigned int irq, int cpu, int mmr_blade,uv_write_global_mmr64 (mmr_pnode, mmr_offset, mmr_value);if (cfg->move_in_progress) -send_cleanup_vector (cfg); +ioapic_send_cleanup_vector (cfg);return irq; }
@@ -208,14 +208,14 @@ static void arch_disable_uv_irq(int mmr_pnode, unsigned long mmr_offset)static int uv_set_irq_affinity (unsigned int irq,const struct cpumask *mask) {struct irq_desc *desc =irq_to_desc (irq); -struct irq_cfg *cfg = desc->chip_data; +struct ioapic_irq_cfg *cfg = desc->chip_data;unsigned int dest;unsigned long mmr_value;struct uv_IO_APIC_route_entry *entry;unsigned long mmr_offset;unsigned mmr_pnode; -if (set_desc_affinity (desc, mask, &dest)) +if (ioapic_set_desc_affinity (desc, mask, &dest))return -1 ; mmr_value =0 ;
@@ -236,7 +236,7 @@ static int uv_set_irq_affinity(unsigned int irq, const struct cpumask *mask)uv_write_global_mmr64 (mmr_pnode, mmr_offset, mmr_value);if (cfg->move_in_progress) -send_cleanup_vector (cfg); +ioapic_send_cleanup_vector (cfg);return 0 ; }
Signature · 1 line
1.5.6.5
Re: [PATCH] x86: namespace some I/O APIC related structures and functions.
Quoted 5 lines · [email protected] writes:
From: Ian Campbell <[email protected]>
It is not obvious that certain functions relate specifically to the I/O APIC. Rename structures and functions as follows:
Quoted 7 lines
struct
irq_cfg-> structioapic_irq_cfgirq_cfg()->ioapic_irq_cfg()assign_irq_vector()->ioapic_assign_irq_vector()send_cleanup_vector()->ioapic_send_cleanup_vector()There is a slight preference towards ioapic vs
io_apicin the current code (563 occurances vs 146) so I went with that.
The rename is wrong. Only the optional irq_2_pin list in struct irq_cfg is specific to ioapics.
The rest of the fields are all about the architectural vector based interrupt delivery system of x86. Branding that structure with ioapic is strongly misleading.
x86_irq_cfg might be an appropriate rename.
Eric
PATCH 1/1 x86: irq_desc->chip_data is always correct whether or not SPARSE_IRQ is enabled.
In reply to Ian CampbellFrom: Ian Campbell <[email protected]>
arch_early_irq_init ensures that in the non-SPARSE_IRQ case that chip_data is only set for irq < NR_IRQS.
--- arch/x86/kernel/apic/io_apic.c | 9 ++------- 1 files changed, 2 insertions(+), 7 deletions(-)
arch/x86/kernel/apic/io_apic.c +2 -7
@@ -178,7 +178,6 @@ int __init arch_early_irq_init(void)return 0 ; } -#ifdef CONFIG_SPARSE_IRQ struct ioapic_irq_cfg *ioapic_irq_cfg (unsigned int irq) {struct ioapic_irq_cfg *cfg =NULL ;
@@ -191,6 +190,8 @@ struct ioapic_irq_cfg *ioapic_irq_cfg(unsigned int irq)return cfg; } +#ifdef CONFIG_SPARSE_IRQ +static struct ioapic_irq_cfg *get_one_free_irq_cfg (int node) {struct ioapic_irq_cfg *cfg;
@@ -333,16 +334,10 @@ void ioapic_free_chip_data(struct irq_desc *old_desc, struct irq_desc *desc)/* end for move_irq_desc */ #else -struct ioapic_irq_cfg *ioapic_irq_cfg (unsigned int irq) -{ -return irq < nr_irqs ? irq_cfgx + irq :NULL ; -} -int ioapic_init_chip_data (struct irq_desc *desc,int node) {return 0 ; } -#endif struct io_apic {
Signature · 1 line
1.5.6.5
Re: [tip:x86/irq] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
In reply to Ian CampbellQuoted 8 lines · Ian Campbell
Quoted 3 lines
- Xen has an array irq_info[NR_IRQS] one of the last static arrays sized at
NR_IRQsin the entire kernel.Hopefully the same info as is in that array could (and indeed should) be instead stored in
irq_desc->chip_data. Would you object toarch_init_copy_chip_dataandarch_free_chip_databecoming function pointers within the structirq_chip?
Yes, I was just about to suggest something like this. We just need a mechanism of storing some per-irq info, and attaching it to the existing desc structure is much cleaner than maintaining some parallel structure.
You're right about the chip_data lifetime functions; it seems very odd that most of irq_chip is nicely factored out into a set of ops, but the per-chip data management is per-architecture.
With that in place, we could eliminate any dependency on irq numbers.
J
[tip:x86/apic] x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
In reply to Brandon PhilipsCommit-ID: 21c2fd9970cc8e457058f84016450a2e9876125e Gitweb: http://git.kernel.org/tip/21c2fd9970cc8e457058f84016450a2e9876125e Author: Ingo Molnar <[email protected]> AuthorDate: Fri, 26 Feb 2010 11:17:16 +0100 Committer: Ingo Molnar <[email protected]> CommitDate: Sat, 27 Feb 2010 12:49:56 +0100
x86: apic: Fix mismerge, add arch_probe_nr_irqs() again
Merge commit aef55d4922 mis-merged io_apic.c so we lost the arch_probe_nr_irqs() method.
This caused subtle boot breakages (udev confusion likely due to missing drivers) with certain configs.
Cc: H. Peter Anvin <[email protected]>
Cc: Yinghai Lu <[email protected]>
LKML-Reference: <[email protected]>
--- arch/x86/kernel/apic/io_apic.c | 22 ++++++++++++++++++++++ 1 files changed, 22 insertions(+), 0 deletions(-)
arch/x86/kernel/apic/io_apic.c +22 -0
@@ -3876,6 +3876,28 @@ void __init probe_nr_irqs_gsi(void)printk (KERN_DEBUG "nr_irqs_gsi: %d\n" , nr_irqs_gsi); } +#ifdef CONFIG_SPARSE_IRQ +int __init arch_probe_nr_irqs (void ) +{ +int nr; + +if (nr_irqs > (NR_VECTORS * nr_cpu_ids)) + nr_irqs =NR_VECTORS * nr_cpu_ids; + + nr = nr_irqs_gsi +8 * nr_cpu_ids; +#if defined(CONFIG_PCI_MSI ) || defined(CONFIG_HT_IRQ ) +/* +* for MSI and HT dyn irq +*/ + nr += nr_irqs_gsi *16 ; +#endif +if (nr < nr_irqs) + nr_irqs = nr; + +return 0 ; +} +#endif +static int __io_apic_set_pci_routing (struct device *dev,int irq,struct io_apic_irq_attr *irq_attr) {
arch/x86/kernel/apic/io_apic.c
Quoted 16 lines · Brandon Philips
Race in
create_irq_nr():- Thread 1 loops through and calls
irq_to_desc_alloc_nodewith new=0x66.- Thread 2 has exited the loop with irq=0x66 and calls
dynamic_irq_init(0x66)settingdesc->chip_data= NULL- Thread 1 then dereferences NULL via
desc_new->chip_data->vectorFix by moving holding
vector_lockuntil after thedynamic_irq_init().Index:
linux-2.6.32-SLE11-SP1/arch/x86/kernel/apic/io_apic.c
===================================================================arch/x86/kernel/apic/io_apic.c +0
-0
can you check if
fix your problem with 2.6.32?
Quoted 1 line
From
37ef2a3029fde884808ff1b369677abc7dd9a79aMon Sep 17 00:00:00 2001
From: Yinghai Lu <[email protected]>
Date: Sat, 21 Nov 2009 00:23:37 -0800
Subject: [PATCH] x86: Re-get cfg_new in case reuse/move irq_desc
When irq_desc is moved, we need to make sure to use the right cfg_new.
LKML-Reference: <[email protected]>
--- arch/x86/kernel/apic/io_apic.c | 1 + 1 files changed, 1 insertions(+), 0 deletions(-)
arch/x86/kernel/apic/io_apic.c +1 -0
@@ -3186,6 +3186,7 @@ unsigned int create_irq_nr(unsigned int irq_want, int node)continue ; desc_new =move_irq_desc (desc_new, node); + cfg_new = desc_new->chip_data;if (__assign_irq_vector (new, cfg_new, apic->target_cpus ()) ==0 ) irq = new;
Signature · 1 line
1.6.6.1