vfio: powerpc/spapr: Move locked_vm accounting to helpers
authorAlexey Kardashevskiy <aik@ozlabs.ru>
Fri, 5 Jun 2015 06:35:01 +0000 (16:35 +1000)
committerMichael Ellerman <mpe@ellerman.id.au>
Thu, 11 Jun 2015 05:14:55 +0000 (15:14 +1000)
There moves locked pages accounting to helpers.
Later they will be reused for Dynamic DMA windows (DDW).

This reworks debug messages to show the current value and the limit.

This stores the locked pages number in the container so when unlocking
the iommu table pointer won't be needed. This does not have an effect
now but it will with the multiple tables per container as then we will
allow attaching/detaching groups on fly and we may end up having
a container with no group attached but with the counter incremented.

While we are here, update the comment explaining why RLIMIT_MEMLOCK
might be required to be bigger than the guest RAM. This also prints
pid of the current process in pr_warn/pr_debug.

Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
[aw: for the vfio related changes]
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
drivers/vfio/vfio_iommu_spapr_tce.c

index 64300ccb05db5596cc40d28a7c7d9f4538ad1726..6e2e15fdb0df9a56bbcb9baa56e9ff6360435f20 100644 (file)
 static void tce_iommu_detach_group(void *iommu_data,
                struct iommu_group *iommu_group);
 
+static long try_increment_locked_vm(long npages)
+{
+       long ret = 0, locked, lock_limit;
+
+       if (!current || !current->mm)
+               return -ESRCH; /* process exited */
+
+       if (!npages)
+               return 0;
+
+       down_write(&current->mm->mmap_sem);
+       locked = current->mm->locked_vm + npages;
+       lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
+       if (locked > lock_limit && !capable(CAP_IPC_LOCK))
+               ret = -ENOMEM;
+       else
+               current->mm->locked_vm += npages;
+
+       pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
+                       npages << PAGE_SHIFT,
+                       current->mm->locked_vm << PAGE_SHIFT,
+                       rlimit(RLIMIT_MEMLOCK),
+                       ret ? " - exceeded" : "");
+
+       up_write(&current->mm->mmap_sem);
+
+       return ret;
+}
+
+static void decrement_locked_vm(long npages)
+{
+       if (!current || !current->mm || !npages)
+               return; /* process exited */
+
+       down_write(&current->mm->mmap_sem);
+       if (WARN_ON_ONCE(npages > current->mm->locked_vm))
+               npages = current->mm->locked_vm;
+       current->mm->locked_vm -= npages;
+       pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
+                       npages << PAGE_SHIFT,
+                       current->mm->locked_vm << PAGE_SHIFT,
+                       rlimit(RLIMIT_MEMLOCK));
+       up_write(&current->mm->mmap_sem);
+}
+
 /*
  * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
  *
@@ -45,6 +90,7 @@ struct tce_container {
        struct mutex lock;
        struct iommu_table *tbl;
        bool enabled;
+       unsigned long locked_pages;
 };
 
 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
@@ -60,7 +106,7 @@ static bool tce_page_is_contained(struct page *page, unsigned page_shift)
 static int tce_iommu_enable(struct tce_container *container)
 {
        int ret = 0;
-       unsigned long locked, lock_limit, npages;
+       unsigned long locked;
        struct iommu_table *tbl = container->tbl;
 
        if (!container->tbl)
@@ -89,21 +135,22 @@ static int tce_iommu_enable(struct tce_container *container)
         * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
         * that would effectively kill the guest at random points, much better
         * enforcing the limit based on the max that the guest can map.
+        *
+        * Unfortunately at the moment it counts whole tables, no matter how
+        * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
+        * each with 2GB DMA window, 8GB will be counted here. The reason for
+        * this is that we cannot tell here the amount of RAM used by the guest
+        * as this information is only available from KVM and VFIO is
+        * KVM agnostic.
         */
-       down_write(&current->mm->mmap_sem);
-       npages = (tbl->it_size << tbl->it_page_shift) >> PAGE_SHIFT;
-       locked = current->mm->locked_vm + npages;
-       lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
-       if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
-               pr_warn("RLIMIT_MEMLOCK (%ld) exceeded\n",
-                               rlimit(RLIMIT_MEMLOCK));
-               ret = -ENOMEM;
-       } else {
+       locked = (tbl->it_size << tbl->it_page_shift) >> PAGE_SHIFT;
+       ret = try_increment_locked_vm(locked);
+       if (ret)
+               return ret;
 
-               current->mm->locked_vm += npages;
-               container->enabled = true;
-       }
-       up_write(&current->mm->mmap_sem);
+       container->locked_pages = locked;
+
+       container->enabled = true;
 
        return ret;
 }
@@ -115,13 +162,10 @@ static void tce_iommu_disable(struct tce_container *container)
 
        container->enabled = false;
 
-       if (!container->tbl || !current->mm)
+       if (!current->mm)
                return;
 
-       down_write(&current->mm->mmap_sem);
-       current->mm->locked_vm -= (container->tbl->it_size <<
-                       container->tbl->it_page_shift) >> PAGE_SHIFT;
-       up_write(&current->mm->mmap_sem);
+       decrement_locked_vm(container->locked_pages);
 }
 
 static void *tce_iommu_open(unsigned long arg)