[rmessiant] 50347: [NTOS] Complete rewrite reserving and releasing of System PTEs. The previous algorithm, in a nutshell, worked as follows: - PTE clusters are in a singly linked list, ordered by t...
9 Jan
2011
9 Jan
'11
8:52 p.m.
Author: rmessiant
Date: Sun Jan 9 20:52:22 2011
New Revision: 50347
URL: http://svn.reactos.org/svn/reactos?rev=50347&view=rev
Log:
[NTOS]
Complete rewrite reserving and releasing of System PTEs.
The previous algorithm, in a nutshell, worked as follows:
- PTE clusters are in a singly linked list, ordered by their base address.
- All PTEs in the clusters are zeroed (except for cluster list bookkeeping).
- Upon reservation: Walk the list to get the first cluster that's large enough, cut
the requested amount of PTEs off its tail and return them.
- Upon release: Create a new cluster using the PTEs to release, and merge it together with
possible adjacent clusters.
Problems with the previous algorithm:
- While the idea is that all PTEs in clusters are zeroed, which requesters rely on,
cluster bookkeeping isn't zeroed on merges.
The side effect of this was that PTEs that weren't really zeroed were randomly
delivered to requesters.
- 99% of the time, allocations are serviced using the first cluster in the list, which is
virtually always the first suitable cluster.
This is so because the ordering is based on the base address of the clusters, and
allocations are serviced using the cluster tail.
Because the first cluster starts out as the whole pool, and the pool is quite sizable,
it can deal with virtually allocations.. for a while.
- A corollary of the previous point is *massive fragmentation* because: as long as an
allocation isn't released back into the pool,
the space of previous allocations that have been released isn't reused because the
first cluster can't suck them up, and enough allocations remain in use.
- The combined effect of the previous two points: a first cluster that effectively shrinks
mostly, with small clusters forming behind it.
Once the first cluster has shrunk far enough (which of course takes a long time), 90% of
the space may still be free, scattered in mostly small clusters.
This would make decent sized allocations fail because of the heavy fragmentation.
- An implementation detail that caused the head of the list to be treated as a genuine
cluster when the first cluster in the list was too small.
The algorithm (as explained above) made this case quite unlikely until your system has
been running for a while, after which it could happily
corrupt list heads of other pools, depending on where the list head is with respect to
its own pool.
Empirically obtained data revealed that after just *booting to the desktop*, the pool for
System Pte Space entries
contained roughly 70 (unusable) clusters, blocking 15 to 20% of the pool. These figures
increased to roughly 100
clusters and 30 to 35% after opening a foxy browser and using it to visit a mathematically
inspired search engine.
The same data also showed that over 95% of allocations requested just a single PTE, and a
noticable allocation spike
also occured in the range of 65-128 PTEs. It should be clear optimizing for small
allocations is a good idea,
and preferably encourage reuse the same PTEs for such allocations.
And the new algorithm was born:
- PTE clusters are in a singly linked list, ordered by increasing cluster size.
- All PTEs in the clusters are zeroed (except for cluster list bookkeeping) .. really this
time!
- Upon reservation: Walk the list to get the first cluster that's large enough, cut
the requested amount of PTEs off its tail and return them.
- Upon release: Create a new cluster using the PTEs to release, and merge it together with
possible adjacent clusters.
- Both in the reservation and release actions, insertions into the list preserve the
increasing cluster size order.
Empirically obtained data now revealed that after just booting to the desktop, the pool
for System Pte Space entries
contained exactly 2 clusters. This increased to 10 clusters after some minor internet
browsing and watching a 5 minute video using a media player.
Modified:
trunk/reactos/ntoskrnl/mm/ARM3/syspte.c
Modified: trunk/reactos/ntoskrnl/mm/ARM3/syspte.c
URL:
http://svn.reactos.org/svn/reactos/trunk/reactos/ntoskrnl/mm/ARM3/syspte.c?…
==============================================================================
--- trunk/reactos/ntoskrnl/mm/ARM3/syspte.c [iso-8859-1] (original)
+++ trunk/reactos/ntoskrnl/mm/ARM3/syspte.c [iso-8859-1] Sun Jan 9 20:52:22 2011
@@ -4,6 +4,7 @@
* FILE: ntoskrnl/mm/ARM3/syspte.c
* PURPOSE: ARM Memory Manager System PTE Allocator
* PROGRAMMERS: ReactOS Portable Systems Group
+ * Roel Messiant (roel.messiant(a)reactos.org)
*/
/* INCLUDES *******************************************************************/
@@ -25,6 +26,55 @@
ULONG MmTotalSystemPtes;
/* PRIVATE FUNCTIONS **********************************************************/
+
+//
+// The free System Page Table Entries are stored in a bunch of clusters,
+// each consisting of one or more PTEs. These PTE clusters are connected
+// in a singly linked list, ordered by increasing cluster size.
+//
+// A cluster consisting of a single PTE is marked by having the OneEntry flag
+// of its PTE set. The forward link is contained in the NextEntry field.
+//
+// Clusters containing multiple PTEs have the OneEntry flag of their first PTE
+// reset. The NextEntry field of the first PTE contains the forward link, and
+// the size of the cluster is stored in the NextEntry field of its second PTE.
+//
+// Reserving PTEs currently happens by walking the linked list until a cluster
+// is found that contains the requested amount of PTEs or more. This cluster
+// is removed from the list, and the requested amount of PTEs is taken from the
+// tail of this cluster. If any PTEs remain in the cluster, the linked list is
+// walked again until a second cluster is found that contains the same amount
+// of PTEs or more. The first cluster is then inserted in front of the second
+// one.
+//
+// Releasing PTEs currently happens by walking the whole linked list, recording
+// the first cluster that contains the amount of PTEs to release or more. When
+// a cluster is found that is adjacent to the PTEs being released, this cluster
+// is removed from the list and subsequently added to the PTEs being released.
+// This ensures no two clusters are adjacent, which maximizes their size.
+// After the walk is complete, a new cluster is created that contains the PTEs
+// being released, which is then inserted in front of the recorded cluster.
+//
+
+/* This definition does not belong here and is most likely platform-dependent */
+#define MM_EMPTY_LIST (ULONG) (0xFFFFF)
+
+ULONG
+FORCEINLINE
+MI_GET_CLUSTER_SIZE(IN PMMPTE Pte)
+{
+ //
+ // First check for a single PTE
+ //
+ if (Pte->u.List.OneEntry)
+ return 1;
+
+ //
+ // Then read the size from the trailing PTE
+ //
+ Pte++;
+ return Pte->u.List.NextEntry;
+}
PMMPTE
NTAPI
@@ -33,8 +83,8 @@
IN ULONG Alignment)
{
KIRQL OldIrql;
- PMMPTE PointerPte, NextPte, PreviousPte;
- ULONG_PTR ClusterSize;
+ PMMPTE PreviousPte, NextPte, ReturnPte;
+ ULONG ClusterSize;
//
// Sanity check
@@ -42,125 +92,146 @@
ASSERT(Alignment <= PAGE_SIZE);
//
- // Lock the system PTE space
+ // Acquire the System PTE lock
//
OldIrql = KeAcquireQueuedSpinLock(LockQueueSystemSpaceLock);
//
- // Get the first free cluster and make sure we have PTEs available
- //
- PointerPte = &MmFirstFreeSystemPte[SystemPtePoolType];
- if (PointerPte->u.List.NextEntry == ((ULONG)0xFFFFF))
- {
- //
- // Fail
+ // Find the last cluster in the list that doesn't contain enough PTEs
+ //
+ PreviousPte = &MmFirstFreeSystemPte[SystemPtePoolType];
+
+ while (PreviousPte->u.List.NextEntry != MM_EMPTY_LIST)
+ {
+ //
+ // Get the next cluster and its size
+ //
+ NextPte = MmSystemPteBase + PreviousPte->u.List.NextEntry;
+ ClusterSize = MI_GET_CLUSTER_SIZE(NextPte);
+
+ //
+ // Check if this cluster contains enough PTEs
+ //
+ if (NumberOfPtes <= ClusterSize)
+ break;
+
+ //
+ // On to the next cluster
+ //
+ PreviousPte = NextPte;
+ }
+
+ //
+ // Make sure we didn't reach the end of the cluster list
+ //
+ if (PreviousPte->u.List.NextEntry == MM_EMPTY_LIST)
+ {
+ //
+ // Release the System PTE lock and return failure
//
KeReleaseQueuedSpinLock(LockQueueSystemSpaceLock, OldIrql);
return NULL;
}
//
- // Now move to the first free system PTE cluster
- //
- PreviousPte = PointerPte;
- PointerPte = MmSystemPteBase + PointerPte->u.List.NextEntry;
-
- //
- // Loop each cluster
- //
- while (TRUE)
- {
- //
- // Check if we're done to only one PTE left
- //
- if (!PointerPte->u.List.OneEntry)
- {
- //
- // Keep track of the next cluster in case we have to relink
- //
- NextPte = PointerPte + 1;
-
- //
- // Can this cluster satisfy the request?
- //
- ClusterSize = (ULONG_PTR)NextPte->u.List.NextEntry;
- if (NumberOfPtes < ClusterSize)
- {
- //
- // It can, and it will leave just one PTE left
- //
- if ((ClusterSize - NumberOfPtes) == 1)
- {
- //
- // This cluster becomes a single system PTE entry
- //
- PointerPte->u.List.OneEntry = 1;
- }
- else
- {
- //
- // Otherwise, the next cluster aborbs what's left
- //
- NextPte->u.List.NextEntry = ClusterSize - NumberOfPtes;
- }
-
- //
- // Decrement the free count and move to the next starting PTE
- //
- MmTotalFreeSystemPtes[SystemPtePoolType] -= NumberOfPtes;
- PointerPte += (ClusterSize - NumberOfPtes);
+ // Unlink the cluster
+ //
+ PreviousPte->u.List.NextEntry = NextPte->u.List.NextEntry;
+
+ //
+ // Check if the reservation spans the whole cluster
+ //
+ if (ClusterSize == NumberOfPtes)
+ {
+ //
+ // Return the first PTE of this cluster
+ //
+ ReturnPte = NextPte;
+
+ //
+ // Zero the cluster
+ //
+ if (NextPte->u.List.OneEntry == 0)
+ {
+ NextPte->u.Long = 0;
+ NextPte++;
+ }
+ NextPte->u.Long = 0;
+ }
+ else
+ {
+ //
+ // Divide the cluster into two parts
+ //
+ ClusterSize -= NumberOfPtes;
+ ReturnPte = NextPte + ClusterSize;
+
+ //
+ // Set the size of the first cluster, zero the second if needed
+ //
+ if (ClusterSize == 1)
+ {
+ NextPte->u.List.OneEntry = 1;
+ ReturnPte->u.Long = 0;
+ }
+ else
+ {
+ NextPte++;
+ NextPte->u.List.NextEntry = ClusterSize;
+ }
+
+ //
+ // Step through the cluster list to find out where to insert the first
+ //
+ PreviousPte = &MmFirstFreeSystemPte[SystemPtePoolType];
+
+ while (PreviousPte->u.List.NextEntry != MM_EMPTY_LIST)
+ {
+ //
+ // Get the next cluster
+ //
+ NextPte = MmSystemPteBase + PreviousPte->u.List.NextEntry;
+
+ //
+ // Check if the cluster to insert is smaller or of equal size
+ //
+ if (ClusterSize <= MI_GET_CLUSTER_SIZE(NextPte))
break;
- }
-
- //
- // Did we find exactly what you wanted?
- //
- if (NumberOfPtes == ClusterSize)
- {
- //
- // Yes, fixup the cluster and decrease free system PTE count
- //
- PreviousPte->u.List.NextEntry = PointerPte->u.List.NextEntry;
- MmTotalFreeSystemPtes[SystemPtePoolType] -= NumberOfPtes;
- break;
- }
- }
- else if (NumberOfPtes == 1)
- {
- //
- // We have one PTE in this cluster, and it's all you want
- //
- PreviousPte->u.List.NextEntry = PointerPte->u.List.NextEntry;
- MmTotalFreeSystemPtes[SystemPtePoolType]--;
- break;
- }
-
- //
- // We couldn't find what you wanted -- is this the last cluster?
- //
- if (PointerPte->u.List.NextEntry == ((ULONG)0xFFFFF))
- {
- //
- // Fail
- //
- KeReleaseQueuedSpinLock(LockQueueSystemSpaceLock, OldIrql);
- return NULL;
- }
-
- //
- // Go to the next cluster
- //
- PreviousPte = PointerPte;
- PointerPte = MmSystemPteBase + PointerPte->u.List.NextEntry;
- ASSERT(PointerPte > PreviousPte);
- }
-
- //
- // Release the lock, flush the TLB and return the first PTE
+
+ //
+ // On to the next cluster
+ //
+ PreviousPte = NextPte;
+ }
+
+ //
+ // Retrieve the first cluster and link it back into the cluster list
+ //
+ NextPte = ReturnPte - ClusterSize;
+
+ NextPte->u.List.NextEntry = PreviousPte->u.List.NextEntry;
+ PreviousPte->u.List.NextEntry = NextPte - MmSystemPteBase;
+ }
+
+ //
+ // Decrease availability
+ //
+ MmTotalFreeSystemPtes[SystemPtePoolType] -= NumberOfPtes;
+
+ //
+ // Release the System PTE lock
//
KeReleaseQueuedSpinLock(LockQueueSystemSpaceLock, OldIrql);
+
+ //
+ // Flush the TLB
+ //
KeFlushProcessTb();
- return PointerPte;
+
+ //
+ // Return the reserved PTEs
+ //
+ return ReturnPte;
}
PMMPTE
@@ -199,24 +270,23 @@
IN MMSYSTEM_PTE_POOL_TYPE SystemPtePoolType)
{
KIRQL OldIrql;
- ULONG_PTR ClusterSize, CurrentSize;
- PMMPTE CurrentPte, NextPte, PointerPte;
+ ULONG_PTR ClusterSize;
+ PMMPTE PreviousPte, NextPte, InsertPte;
//
// Check to make sure the PTE address is within bounds
//
ASSERT(NumberOfPtes != 0);
ASSERT(StartingPte >= MmSystemPtesStart[SystemPtePoolType]);
- ASSERT(StartingPte <= MmSystemPtesEnd[SystemPtePoolType]);
+ ASSERT(StartingPte + NumberOfPtes - 1 <= MmSystemPtesEnd[SystemPtePoolType]);
//
// Zero PTEs
//
RtlZeroMemory(StartingPte, NumberOfPtes * sizeof(MMPTE));
- CurrentSize = (ULONG_PTR)(StartingPte - MmSystemPteBase);
-
- //
- // Acquire the system PTE lock
+
+ //
+ // Acquire the System PTE lock
//
OldIrql = KeAcquireQueuedSpinLock(LockQueueSystemSpaceLock);
@@ -226,141 +296,94 @@
MmTotalFreeSystemPtes[SystemPtePoolType] += NumberOfPtes;
//
- // Get the free cluster and start going through them
- //
- CurrentPte = &MmFirstFreeSystemPte[SystemPtePoolType];
- while (TRUE)
- {
- //
- // Get the first real cluster of PTEs and check if it's ours
- //
- PointerPte = MmSystemPteBase + CurrentPte->u.List.NextEntry;
- if (CurrentSize < CurrentPte->u.List.NextEntry)
- {
- //
- // Sanity check
- //
- ASSERT(((StartingPte + NumberOfPtes) <= PointerPte) ||
- (CurrentPte->u.List.NextEntry == ((ULONG)0xFFFFF)));
-
- //
- // Get the next cluster in case it's the one
- //
- NextPte = CurrentPte + 1;
-
- //
- // Check if this was actually a single-PTE entry
- //
- if (CurrentPte->u.List.OneEntry)
+ // Step through the cluster list to find where to insert the PTEs
+ //
+ PreviousPte = &MmFirstFreeSystemPte[SystemPtePoolType];
+ InsertPte = NULL;
+
+ while (PreviousPte->u.List.NextEntry != MM_EMPTY_LIST)
+ {
+ //
+ // Get the next cluster and its size
+ //
+ NextPte = MmSystemPteBase + PreviousPte->u.List.NextEntry;
+ ClusterSize = MI_GET_CLUSTER_SIZE(NextPte);
+
+ //
+ // Check if this cluster is adjacent to the PTEs being released
+ //
+ if ((NextPte + ClusterSize == StartingPte) ||
+ (StartingPte + NumberOfPtes == NextPte))
+ {
+ //
+ // Add the PTEs in the cluster to the PTEs being released
+ //
+ NumberOfPtes += ClusterSize;
+
+ if (NextPte < StartingPte)
+ StartingPte = NextPte;
+
+ //
+ // Unlink this cluster and zero it
+ //
+ PreviousPte->u.List.NextEntry = NextPte->u.List.NextEntry;
+
+ if (NextPte->u.List.OneEntry == 0)
{
- //
- // We only have one page
- //
- ClusterSize = 1;
+ NextPte->u.Long = 0;
+ NextPte++;
}
- else
- {
- //
- // The next cluster will have the page count
- //
- ClusterSize = (ULONG_PTR)NextPte->u.List.NextEntry;
- }
-
- //
- // So check if this cluster actually describes the entire mapping
- //
- if ((CurrentPte + ClusterSize) == StartingPte)
- {
- //
- // It does -- collapse the free PTEs into the next cluster
- //
- NumberOfPtes += ClusterSize;
- NextPte->u.List.NextEntry = NumberOfPtes;
- CurrentPte->u.List.OneEntry = 0;
-
- //
- // Make another pass
- //
- StartingPte = CurrentPte;
- }
- else
- {
- //
- // There's still PTEs left -- make us into a cluster
- //
- StartingPte->u.List.NextEntry = CurrentPte->u.List.NextEntry;
- CurrentPte->u.List.NextEntry = CurrentSize;
-
- //
- // Is there just one page left?
- //
- if (NumberOfPtes == 1)
- {
- //
- // Then this actually becomes a single PTE entry
- //
- StartingPte->u.List.OneEntry = 1;
- }
- else
- {
- //
- // Otherwise, create a new cluster for the remaining pages
- //
- StartingPte->u.List.OneEntry = 0;
- NextPte = StartingPte + 1;
- NextPte->u.List.NextEntry = NumberOfPtes;
- }
- }
-
- //
- // Now check if we've arrived at yet another cluster
- //
- if ((StartingPte + NumberOfPtes) == PointerPte)
- {
- //
- // We'll collapse the next cluster into us
- //
- StartingPte->u.List.NextEntry = PointerPte->u.List.NextEntry;
- StartingPte->u.List.OneEntry = 0;
- NextPte = StartingPte + 1;
-
- //
- // Check if the cluster only had one page
- //
- if (PointerPte->u.List.OneEntry)
- {
- //
- // So will we...
- //
- ClusterSize = 1;
- }
- else
- {
- //
- // Otherwise, grab the page count from the next-next cluster
- //
- PointerPte++;
- ClusterSize = (ULONG_PTR)PointerPte->u.List.NextEntry;
- }
-
- //
- // And create the final combined cluster
- //
- NextPte->u.List.NextEntry = NumberOfPtes + ClusterSize;
- }
-
- //
- // We released the PTEs into their cluster (and optimized the list)
- //
- KeReleaseQueuedSpinLock(LockQueueSystemSpaceLock, OldIrql);
- break;
- }
-
- //
- // Try the next cluster of PTEs...
- //
- CurrentPte = PointerPte;
- }
+ NextPte->u.Long = 0;
+
+ //
+ // Invalidate the previously found insertion location, if any
+ //
+ InsertPte = NULL;
+ }
+ else
+ {
+ //
+ // Check if the insertion location is right before this cluster
+ //
+ if ((InsertPte == NULL) && (NumberOfPtes <= ClusterSize))
+ InsertPte = PreviousPte;
+
+ //
+ // On to the next cluster
+ //
+ PreviousPte = NextPte;
+ }
+ }
+
+ //
+ // If no insertion location was found, use the tail of the list
+ //
+ if (InsertPte == NULL)
+ InsertPte = PreviousPte;
+
+ //
+ // Create a new cluster using the PTEs being released
+ //
+ if (NumberOfPtes != 1)
+ {
+ StartingPte->u.List.OneEntry = 0;
+
+ NextPte = StartingPte + 1;
+ NextPte->u.List.NextEntry = NumberOfPtes;
+ }
+ else
+ StartingPte->u.List.OneEntry = 1;
+
+ //
+ // Link the new cluster into the cluster list at the insertion location
+ //
+ StartingPte->u.List.NextEntry = InsertPte->u.List.NextEntry;
+ InsertPte->u.List.NextEntry = StartingPte - MmSystemPteBase;
+
+ //
+ // Release the System PTE lock
+ //
+ KeReleaseQueuedSpinLock(LockQueueSystemSpaceLock, OldIrql);
}
VOID
@@ -392,7 +415,7 @@
//
// Make the first entry free and link it
//
- StartingPte->u.List.NextEntry = ((ULONG)0xFFFFF);
+ StartingPte->u.List.NextEntry = MM_EMPTY_LIST;
MmFirstFreeSystemPte[PoolType].u.Long = 0;
MmFirstFreeSystemPte[PoolType].u.List.NextEntry = StartingPte -
MmSystemPteBase;
5003
days inactive
5003
days old
0 comments
1 participants
participants (1)
-
rmessiant＠svn.reactos.org