[PATCH 5/5] powerpc: iommu: Implement IOMMU pools to improve multiqueue adapter performance
Anton Blanchard
anton at samba.org
Fri Jun 8 14:14:48 EST 2012
At the moment all queues in a multiqueue adapter will serialise
against the IOMMU table lock. This is proving to be a big issue,
especially with 10Gbit ethernet.
This patch creates 4 pools and tries to spread the load across
them. If the table is under 1GB in size we revert back to the
original behaviour of 1 pool and 1 largealloc pool.
We create a hash to map CPUs to pools. Since we prefer interrupts to
be affinitised to primary CPUs, without some form of hashing we are
very likely to end up using the same pool. As an example, POWER7
has 4 way SMT and with 4 pools all primary threads will map to the
same pool.
The largealloc pool is reduced from 1/2 to 1/4 of the space to
partially offset the overhead of breaking the table up into pools.
Some performance numbers were obtained with a Chelsio T3 adapter on
two POWER7 boxes, running a 100 session TCP round robin test.
Performance improved 69% with this patch applied.
Signed-off-by: Anton Blanchard <anton at samba.org>
---
All patches combined improve performance by 178%
v2.0: Fix cell build, noticed by mpe.
Index: linux-build/arch/powerpc/kernel/iommu.c
===================================================================
--- linux-build.orig/arch/powerpc/kernel/iommu.c 2012-06-08 14:02:46.229104320 +1000
+++ linux-build/arch/powerpc/kernel/iommu.c 2012-06-08 14:02:46.861113517 +1000
@@ -33,6 +33,7 @@
#include <linux/bitmap.h>
#include <linux/iommu-helper.h>
#include <linux/crash_dump.h>
+#include <linux/hash.h>
#include <asm/io.h>
#include <asm/prom.h>
#include <asm/iommu.h>
@@ -58,6 +59,26 @@ static int __init setup_iommu(char *str)
__setup("iommu=", setup_iommu);
+static DEFINE_PER_CPU(unsigned int, iommu_pool_hash);
+
+/*
+ * We precalculate the hash to avoid doing it on every allocation.
+ *
+ * The hash is important to spread CPUs across all the pools. For example,
+ * on a POWER7 with 4 way SMT we want interrupts on the primary threads and
+ * with 4 pools all primary threads would map to the same pool.
+ */
+static int __init setup_iommu_pool_hash(void)
+{
+ unsigned int i;
+
+ for_each_possible_cpu(i)
+ per_cpu(iommu_pool_hash, i) = hash_32(i, IOMMU_POOL_HASHBITS);
+
+ return 0;
+}
+subsys_initcall(setup_iommu_pool_hash);
+
static unsigned long iommu_range_alloc(struct device *dev,
struct iommu_table *tbl,
unsigned long npages,
@@ -72,6 +93,8 @@ static unsigned long iommu_range_alloc(s
unsigned long align_mask;
unsigned long boundary_size;
unsigned long flags;
+ unsigned int pool_nr;
+ struct iommu_pool *pool;
align_mask = 0xffffffffffffffffl >> (64 - align_order);
@@ -84,38 +107,46 @@ static unsigned long iommu_range_alloc(s
return DMA_ERROR_CODE;
}
- spin_lock_irqsave(&(tbl->it_lock), flags);
+ /*
+ * We don't need to disable preemption here because any CPU can
+ * safely use any IOMMU pool.
+ */
+ pool_nr = __raw_get_cpu_var(iommu_pool_hash) & (tbl->nr_pools - 1);
- if (handle && *handle)
- start = *handle;
+ if (largealloc)
+ pool = &(tbl->large_pool);
else
- start = largealloc ? tbl->it_largehint : tbl->it_hint;
+ pool = &(tbl->pools[pool_nr]);
+
+ spin_lock_irqsave(&(pool->lock), flags);
- /* Use only half of the table for small allocs (15 pages or less) */
- limit = largealloc ? tbl->it_size : tbl->it_halfpoint;
+again:
+ if ((pass == 0) && handle && *handle)
+ start = *handle;
+ else
+ start = pool->hint;
- if (largealloc && start < tbl->it_halfpoint)
- start = tbl->it_halfpoint;
+ limit = pool->end;
/* The case below can happen if we have a small segment appended
* to a large, or when the previous alloc was at the very end of
* the available space. If so, go back to the initial start.
*/
if (start >= limit)
- start = largealloc ? tbl->it_largehint : tbl->it_hint;
-
- again:
+ start = pool->start;
if (limit + tbl->it_offset > mask) {
limit = mask - tbl->it_offset + 1;
/* If we're constrained on address range, first try
* at the masked hint to avoid O(n) search complexity,
- * but on second pass, start at 0.
+ * but on second pass, start at 0 in pool 0.
*/
- if ((start & mask) >= limit || pass > 0)
- start = 0;
- else
+ if ((start & mask) >= limit || pass > 0) {
+ pool = &(tbl->pools[0]);
+ start = pool->start;
+ } else {
start &= mask;
+ }
}
if (dev)
@@ -129,17 +160,25 @@ static unsigned long iommu_range_alloc(s
tbl->it_offset, boundary_size >> IOMMU_PAGE_SHIFT,
align_mask);
if (n == -1) {
- if (likely(pass < 2)) {
- /* First failure, just rescan the half of the table.
- * Second failure, rescan the other half of the table.
- */
- start = (largealloc ^ pass) ? tbl->it_halfpoint : 0;
- limit = pass ? tbl->it_size : limit;
+ if (likely(pass == 0)) {
+ /* First try the pool from the start */
+ pool->hint = pool->start;
pass++;
goto again;
+
+ } else if (pass <= tbl->nr_pools) {
+ /* Now try scanning all the other pools */
+ spin_unlock(&(pool->lock));
+ pool_nr = (pool_nr + 1) & (tbl->nr_pools - 1);
+ pool = &tbl->pools[pool_nr];
+ spin_lock(&(pool->lock));
+ pool->hint = pool->start;
+ pass++;
+ goto again;
+
} else {
- /* Third failure, give up */
- spin_unlock_irqrestore(&(tbl->it_lock), flags);
+ /* Give up */
+ spin_unlock_irqrestore(&(pool->lock), flags);
return DMA_ERROR_CODE;
}
}
@@ -149,10 +188,10 @@ static unsigned long iommu_range_alloc(s
/* Bump the hint to a new block for small allocs. */
if (largealloc) {
/* Don't bump to new block to avoid fragmentation */
- tbl->it_largehint = end;
+ pool->hint = end;
} else {
/* Overflow will be taken care of at the next allocation */
- tbl->it_hint = (end + tbl->it_blocksize - 1) &
+ pool->hint = (end + tbl->it_blocksize - 1) &
~(tbl->it_blocksize - 1);
}
@@ -160,7 +199,8 @@ static unsigned long iommu_range_alloc(s
if (handle)
*handle = end;
- spin_unlock_irqrestore(&(tbl->it_lock), flags);
+ spin_unlock_irqrestore(&(pool->lock), flags);
+
return n;
}
@@ -235,23 +275,45 @@ static bool iommu_free_check(struct iomm
return true;
}
+static struct iommu_pool *get_pool(struct iommu_table *tbl,
+ unsigned long entry)
+{
+ struct iommu_pool *p;
+ unsigned long largepool_start = tbl->large_pool.start;
+
+ /* The large pool is the last pool at the top of the table */
+ if (entry >= largepool_start) {
+ p = &tbl->large_pool;
+ } else {
+ unsigned int pool_nr = entry / tbl->poolsize;
+
+ BUG_ON(pool_nr > tbl->nr_pools);
+ p = &tbl->pools[pool_nr];
+ }
+
+ return p;
+}
+
static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
unsigned int npages)
{
unsigned long entry, free_entry;
unsigned long flags;
+ struct iommu_pool *pool;
entry = dma_addr >> IOMMU_PAGE_SHIFT;
free_entry = entry - tbl->it_offset;
+ pool = get_pool(tbl, free_entry);
+
if (!iommu_free_check(tbl, dma_addr, npages))
return;
ppc_md.tce_free(tbl, entry, npages);
- spin_lock_irqsave(&(tbl->it_lock), flags);
+ spin_lock_irqsave(&(pool->lock), flags);
bitmap_clear(tbl->it_map, free_entry, npages);
- spin_unlock_irqrestore(&(tbl->it_lock), flags);
+ spin_unlock_irqrestore(&(pool->lock), flags);
}
static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
@@ -493,9 +555,8 @@ struct iommu_table *iommu_init_table(str
unsigned long sz;
static int welcomed = 0;
struct page *page;
-
- /* Set aside 1/4 of the table for large allocations. */
- tbl->it_halfpoint = tbl->it_size * 3 / 4;
+ unsigned int i;
+ struct iommu_pool *p;
/* number of bytes needed for the bitmap */
sz = (tbl->it_size + 7) >> 3;
@@ -514,9 +575,28 @@ struct iommu_table *iommu_init_table(str
if (tbl->it_offset == 0)
set_bit(0, tbl->it_map);
- tbl->it_hint = 0;
- tbl->it_largehint = tbl->it_halfpoint;
- spin_lock_init(&tbl->it_lock);
+ /* We only split the IOMMU table if we have 1GB or more of space */
+ if ((tbl->it_size << IOMMU_PAGE_SHIFT) >= (1UL * 1024 * 1024 * 1024))
+ tbl->nr_pools = IOMMU_NR_POOLS;
+ else
+ tbl->nr_pools = 1;
+
+ /* We reserve the top 1/4 of the table for large allocations */
+ tbl->poolsize = (tbl->it_size * 3 / 4) / IOMMU_NR_POOLS;
+
+ for (i = 0; i < IOMMU_NR_POOLS; i++) {
+ p = &tbl->pools[i];
+ spin_lock_init(&(p->lock));
+ p->start = tbl->poolsize * i;
+ p->hint = p->start;
+ p->end = p->start + tbl->poolsize;
+ }
+
+ p = &tbl->large_pool;
+ spin_lock_init(&(p->lock));
+ p->start = tbl->poolsize * i;
+ p->hint = p->start;
+ p->end = tbl->it_size;
iommu_table_clear(tbl);
Index: linux-build/arch/powerpc/include/asm/iommu.h
===================================================================
--- linux-build.orig/arch/powerpc/include/asm/iommu.h 2012-06-06 14:44:13.106458136 +1000
+++ linux-build/arch/powerpc/include/asm/iommu.h 2012-06-08 14:02:46.861113517 +1000
@@ -53,6 +53,16 @@ static __inline__ __attribute_const__ in
*/
#define IOMAP_MAX_ORDER 13
+#define IOMMU_POOL_HASHBITS 2
+#define IOMMU_NR_POOLS (1 << IOMMU_POOL_HASHBITS)
+
+struct iommu_pool {
+ unsigned long start;
+ unsigned long end;
+ unsigned long hint;
+ spinlock_t lock;
+} ____cacheline_aligned_in_smp;
+
struct iommu_table {
unsigned long it_busno; /* Bus number this table belongs to */
unsigned long it_size; /* Size of iommu table in entries */
@@ -61,10 +71,10 @@ struct iommu_table {
unsigned long it_index; /* which iommu table this is */
unsigned long it_type; /* type: PCI or Virtual Bus */
unsigned long it_blocksize; /* Entries in each block (cacheline) */
- unsigned long it_hint; /* Hint for next alloc */
- unsigned long it_largehint; /* Hint for large allocs */
- unsigned long it_halfpoint; /* Breaking point for small/large allocs */
- spinlock_t it_lock; /* Protects it_map */
+ unsigned long poolsize;
+ unsigned long nr_pools;
+ struct iommu_pool large_pool;
+ struct iommu_pool pools[IOMMU_NR_POOLS];
unsigned long *it_map; /* A simple allocation bitmap for now */
};
Index: linux-build/arch/powerpc/platforms/cell/iommu.c
===================================================================
--- linux-build.orig/arch/powerpc/platforms/cell/iommu.c 2012-04-05 13:47:45.715857539 +1000
+++ linux-build/arch/powerpc/platforms/cell/iommu.c 2012-06-08 14:03:00.053305540 +1000
@@ -518,7 +518,6 @@ cell_iommu_setup_window(struct cbe_iommu
__set_bit(0, window->table.it_map);
tce_build_cell(&window->table, window->table.it_offset, 1,
(unsigned long)iommu->pad_page, DMA_TO_DEVICE, NULL);
- window->table.it_hint = window->table.it_blocksize;
return window;
}
More information about the Linuxppc-dev
mailing list