Em Thu, May 15, 2008 at 01:04:26PM -0300, Arnaldo Carvalho de Melo escreveu: > Hi, > > While using the preemptirqsoff ftrace tracer I noticed that > everytime handle_edge_irq is called it needs to mask and unmask MSI, and > that leads to a series of very expensive calls to pci_find_capability, > as can be seen here, with preemption disabled: > > <idle>-0 [03] 422.558652: unmask_msi_irq <-handle_edge_irq > <idle>-0 [03] 422.558653: msi_set_mask_bits <-unmask_msi_irq > <idle>-0 [03] 422.558653: msi_set_enable <-msi_set_mask_bits > <idle>-0 [03] 422.558654: pci_find_capability <-msi_set_enable > <idle>-0 [03] 422.558655: __pci_bus_find_cap_start <-pci_find_capability > <idle>-0 [03] 422.558655: pci_bus_read_config_word <-__pci_bus_find_cap_start > <idle>-0 [03] 422.558656: _spin_lock_irqsave <-pci_bus_read_config_word > <idle>-0 [03] 422.558656: add_preempt_count <-_spin_lock_irqsave > <idle>-0 [03] 422.558657: pci_read <-pci_bus_read_config_word > <idle>-0 [03] 422.558657: raw_pci_read <-pci_read > <idle>-0 [03] 422.558658: pci_conf1_read <-raw_pci_read > <idle>-0 [03] 422.558658: _spin_lock_irqsave <-pci_conf1_read > <idle>-0 [03] 422.558659: add_preempt_count <-_spin_lock_irqsave > > BZZT! 37us > > <idle>-0 [03] 422.558696: _spin_unlock_irqrestore <-pci_conf1_read > <idle>-0 [03] 422.558697: sub_preempt_count <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558697: preempt_schedule <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558698: _spin_unlock_irqrestore <-pci_bus_read_config_word > <idle>-0 [03] 422.558698: sub_preempt_count <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558699: preempt_schedule <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558699: __pci_find_next_cap <-pci_find_capability > <idle>-0 [03] 422.558700: __pci_find_next_cap_ttl <-__pci_find_next_cap > <idle>-0 [03] 422.558700: pci_bus_read_config_byte <-__pci_find_next_cap_ttl > <idle>-0 [03] 422.558701: _spin_lock_irqsave <-pci_bus_read_config_byte > <idle>-0 [03] 422.558701: add_preempt_count <-_spin_lock_irqsave > <idle>-0 [03] 422.558702: pci_read <-pci_bus_read_config_byte > <idle>-0 [03] 422.558702: raw_pci_read <-pci_read > <idle>-0 [03] 422.558703: pci_conf1_read <-raw_pci_read > <idle>-0 [03] 422.558703: _spin_lock_irqsave <-pci_conf1_read > <idle>-0 [03] 422.558704: add_preempt_count <-_spin_lock_irqsave > > BZZT! 38us > > <idle>-0 [03] 422.558742: _spin_unlock_irqrestore <-pci_conf1_read > <idle>-0 [03] 422.558743: sub_preempt_count <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558743: preempt_schedule <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558744: _spin_unlock_irqrestore <-pci_bus_read_config_byte > <idle>-0 [03] 422.558744: sub_preempt_count <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558745: preempt_schedule <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558745: pci_bus_read_config_byte <-__pci_find_next_cap_ttl > <idle>-0 [03] 422.558746: _spin_lock_irqsave <-pci_bus_read_config_byte > <idle>-0 [03] 422.558746: add_preempt_count <-_spin_lock_irqsave > <idle>-0 [03] 422.558747: pci_read <-pci_bus_read_config_byte > <idle>-0 [03] 422.558747: raw_pci_read <-pci_read > <idle>-0 [03] 422.558748: pci_conf1_read <-raw_pci_read > <idle>-0 [03] 422.558748: _spin_lock_irqsave <-pci_conf1_read > <idle>-0 [03] 422.558749: add_preempt_count <-_spin_lock_irqsave > > BZZT! 39us > > <idle>-0 [03] 422.558788: _spin_unlock_irqrestore <-pci_conf1_read > <idle>-0 [03] 422.558789: sub_preempt_count <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558789: preempt_schedule <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558790: _spin_unlock_irqrestore <-pci_bus_read_config_byte > <idle>-0 [03] 422.558790: sub_preempt_count <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558791: preempt_schedule <-_spin_unlock_irqrestore > <idle>-0 [03] 422.558791: pci_bus_read_config_byte <-__pci_find_next_cap_ttl > <idle>-0 [03] 422.558792: _spin_lock_irqsave <-pci_bus_read_config_byte > <idle>-0 [03] 422.558792: add_preempt_count <-_spin_lock_irqsave > <idle>-0 [03] 422.558793: pci_read <-pci_bus_read_config_byte > <idle>-0 [03] 422.558793: raw_pci_read <-pci_read > <idle>-0 [03] 422.558794: pci_conf1_read <-raw_pci_read > <idle>-0 [03] 422.558794: _spin_lock_irqsave <-pci_conf1_read > <idle>-0 [03] 422.558795: add_preempt_count <-_spin_lock_irqsave > > BZZT! 39us > > <idle>-0 [03] 422.558834: _spin_unlock_irqrestore <-pci_conf1_read > <idle>-0 [03] 422.558834: sub_preempt_count <-_spin_unlock_irqrestore > > <SNIP many more such BZZT!s> > > So I implemented pci_find_capability_cached and made MSI use it > for good measure, please consider applying. > > Signed-off-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> > > diff --git a/Makefile b/Makefile > index 14f34b4..d79fdac 100644 > --- a/Makefile > +++ b/Makefile > @@ -1,7 +1,7 @@ > VERSION = 2 > PATCHLEVEL = 6 > SUBLEVEL = 25 > -EXTRAVERSION = > +EXTRAVERSION = .pci_cached > NAME = Funky Weasel is Jiggy wit it > > # *DOCUMENTATION* Ouch, left this in, do you want another patch or can you just remove this bit? Thanks, - Arnaldo -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
