On Thu, May 3, 2012 at 12:13 AM, Jiang Liu [off-list ref] wrote:

Hi Bjorn,
       Thanks for your comments!
       I have thought about using pci_config_lock to serialize
access to the pci_mmcfg_list, but that may hurt MMCFG on x86_64
systems because mmconfig_64.c supports concurrent access to
the MMCFG space without holding the pci_config_lock. So I use
RCU to avoid possible performance penalty for x86_64 systems.

Oh, right.  I only looked at mmconfig_32.c, which uses the lock, and
assumed _64.c was the same.  I think you're right to avoid introducing
locking where we didn't have it before.

       There are two mechanisms to protect list_add_sorted()
from concurrent updates. The first case is, list_add_sorted()
may be called without holding any lock for serialization at
early booting stages because the system is still in single-threaded
mode. The second case is, pci_config_lock is used to serialize
concurrent modification to the pci_mmcfg_list if list_add_sorted()
is called at runtime.
       The first case is cover by this patch, and the second case
is covered by "[PATCH V2 5/6] PCI, x86: introduce
pci_mmconfig_insert()/delete() for PCI root bridge hotplug".

How hard would it be to convert the first (early boot) mechanism to
use the second mechanism?  It'd be nicer if there were only one path
that we use both places, even if it means the early boot path acquires
a lock that is technically unnecessary.

On 2012-5-3 6:31, Bjorn Helgaas wrote:

quoted

On Mon, Apr 9, 2012 at 10:22 AM, Jiang Liu[off-list ref]  wrote:

quoted

Use RCU list to protect mmconfig list from dynamic change
when supporting PCI host bridge hotplug.

I'm not convinced this is safe.  But I'm not an RCU expert, so maybe
you can convince me :)

Documentation/RCU/checklist.txt says the update code still requires
mutual exclusion, and I don't see any in list_add_sorted() or its
callers.

MMCONFIG accesses are already protected by the pci_config_lock
spinlock, so I don't think the performance advantages of RCU really
gain us anything in this situation.

It would definitely be simpler to just hold pci_config_lock while
updating and searching the pci_mmcfg_list, but if there are issues
with that, you can educate me about what they are and why this RCU
code is correct.

Bjorn

quoted

Signed-off-by: Jiang Liu<redacted>
---
 arch/x86/pci/mmconfig-shared.c |   11 ++++++-----
 arch/x86/pci/mmconfig_32.c     |   13 +++++++++++--
 arch/x86/pci/mmconfig_64.c     |   13 +++++++++++--
 3 files changed, 28 insertions(+), 9 deletions(-)

diff --git a/arch/x86/pci/mmconfig-shared.c

b/arch/x86/pci/mmconfig-shared.c
index 5e2cd2a..3bcc361 100644

--- a/arch/x86/pci/mmconfig-shared.c
+++ b/arch/x86/pci/mmconfig-shared.c

@@ -17,6 +17,7 @@

 #include<linux/bitmap.h>
 #include<linux/dmi.h>
 #include<linux/slab.h>
+#include<linux/rculist.h>
 #include<asm/e820.h>
 #include<asm/pci_x86.h>
 #include<asm/acpi.h>

@@ -45,20 +46,20 @@ static __init void free_all_mmcfg(void)

               pci_mmconfig_remove(cfg);
 }

-static __init void list_add_sorted(struct pci_mmcfg_region *new)
+static __devinit void list_add_sorted(struct pci_mmcfg_region *new)
 {
       struct pci_mmcfg_region *cfg;

       /* keep list sorted by segment and starting bus number */
-       list_for_each_entry(cfg,&pci_mmcfg_list, list) {
+       list_for_each_entry_rcu(cfg,&pci_mmcfg_list, list) {

               if (cfg->segment>  new->segment ||
                   (cfg->segment == new->segment&&
                    cfg->start_bus>= new->start_bus)) {
-                       list_add_tail(&new->list,&cfg->list);
+                       list_add_tail_rcu(&new->list,&cfg->list);
                       return;
               }
       }
-       list_add_tail(&new->list,&pci_mmcfg_list);
+       list_add_tail_rcu(&new->list,&pci_mmcfg_list);
 }

 static __devinit struct pci_mmcfg_region *pci_mmconfig_alloc(int
segment,

@@ -111,7 +112,7 @@ struct pci_mmcfg_region *pci_mmconfig_lookup(int

segment, int bus)
 {
       struct pci_mmcfg_region *cfg;

-       list_for_each_entry(cfg,&pci_mmcfg_list, list)
+       list_for_each_entry_rcu(cfg,&pci_mmcfg_list, list)

               if (cfg->segment == segment&&
                   cfg->start_bus<= bus&&  bus<= cfg->end_bus)

                       return cfg;

diff --git a/arch/x86/pci/mmconfig_32.c b/arch/x86/pci/mmconfig_32.c
index 5372e86..5dad04a 100644
--- a/arch/x86/pci/mmconfig_32.c
+++ b/arch/x86/pci/mmconfig_32.c

@@ -11,6 +11,7 @@

 #include<linux/pci.h>
 #include<linux/init.h>
+#include<linux/rcupdate.h>
 #include<asm/e820.h>
 #include<asm/pci_x86.h>
 #include<acpi/acpi.h>

@@ -60,9 +61,12 @@ err:         *value = -1;

               return -EINVAL;
       }

+       rcu_read_lock();
       base = get_base_addr(seg, bus, devfn);
-       if (!base)
+       if (!base) {
+               rcu_read_unlock();
               goto err;
+       }

       raw_spin_lock_irqsave(&pci_config_lock, flags);

@@ -80,6 +84,7 @@ err:          *value = -1;

               break;
       }
       raw_spin_unlock_irqrestore(&pci_config_lock, flags);
+       rcu_read_unlock();

       return 0;
 }

@@ -93,9 +98,12 @@ static int pci_mmcfg_write(unsigned int seg, unsigned

int bus,
       if ((bus>  255) || (devfn>  255) || (reg>  4095))
               return -EINVAL;

+       rcu_read_lock();
       base = get_base_addr(seg, bus, devfn);
-       if (!base)
+       if (!base) {
+               rcu_read_unlock();
               return -EINVAL;
+       }

       raw_spin_lock_irqsave(&pci_config_lock, flags);

@@ -113,6 +121,7 @@ static int pci_mmcfg_write(unsigned int seg, unsigned

int bus,
               break;
       }
       raw_spin_unlock_irqrestore(&pci_config_lock, flags);
+       rcu_read_unlock();

       return 0;
 }

diff --git a/arch/x86/pci/mmconfig_64.c b/arch/x86/pci/mmconfig_64.c
index 915a493..acc48c5 100644
--- a/arch/x86/pci/mmconfig_64.c
+++ b/arch/x86/pci/mmconfig_64.c

@@ -9,6 +9,7 @@

 #include<linux/init.h>
 #include<linux/acpi.h>
 #include<linux/bitmap.h>
+#include<linux/rcupdate.h>
 #include<asm/e820.h>
 #include<asm/pci_x86.h>

@@ -34,9 +35,12 @@ err:         *value = -1;

               return -EINVAL;
       }

+       rcu_read_lock();
       addr = pci_dev_base(seg, bus, devfn);
-       if (!addr)
+       if (!addr) {
+               rcu_read_unlock();
               goto err;
+       }

       switch (len) {
       case 1:

@@ -49,6 +53,7 @@ err:          *value = -1;

               *value = mmio_config_readl(addr + reg);
               break;
       }
+       rcu_read_unlock();

       return 0;
 }

@@ -62,9 +67,12 @@ static int pci_mmcfg_write(unsigned int seg, unsigned

int bus,
       if (unlikely((bus>  255) || (devfn>  255) || (reg>  4095)))
               return -EINVAL;

+       rcu_read_lock();
       addr = pci_dev_base(seg, bus, devfn);
-       if (!addr)
+       if (!addr) {
+               rcu_read_unlock();
               return -EINVAL;
+       }

       switch (len) {
       case 1:

@@ -77,6 +85,7 @@ static int pci_mmcfg_write(unsigned int seg, unsigned

int bus,
               mmio_config_writel(addr + reg, value);
               break;
       }
+       rcu_read_unlock();

       return 0;
 }
--
1.7.5.4

.

--
To unsubscribe from this list: send the line "unsubscribe linux-acpi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html