Refine slab cache sizing

This change is designed to improve the memory utilization of
slabs by more carefully setting their size.  The way the code
currently works is problematic for slabs which contain large
objects (>1MB).  This is due to slabs being unconditionally
rounded up to a power of two which may result in unused space
at the end of the slab.

The reason the existing code rounds up every slab is because it
assumes it will backed by the buddy allocator.  Since the buddy
allocator can only performs power of two allocations this is
desirable because it avoids wasting any space.  However, this
logic breaks down if slab is backed by vmalloc() which operates
at a page level granularity.  In this case, the optimal thing to
do is calculate the minimum required slab size given certain
constraints (object size, alignment, objects/slab, etc).

Therefore, this patch reworks the spl_slab_size() function so
that it sizes KMC_KMEM slabs differently than KMC_VMEM slabs.
KMC_KMEM slabs are rounded up to the nearest power of two, and
KMC_VMEM slabs are allowed to be the minimum required size.

This change also reduces the default number of objects per slab.
This reduces how much memory a single cache object can pin, which
can result in significant memory saving for highly fragmented
caches.  But depending on the workload it may result in slabs
being allocated and freed more frequently.  In practice, this
has been shown to be a better default for most workloads.

Also the maximum slab size has been reduced to 4MB on 32-bit
systems.  Due to the limited virtual address space it's critical
the we be as frugal as possible.  A limit of 4M still lets us
reasonably comfortably allocate a limited number of 1MB objects.

Finally, the kmem:slab_small and kmem:slab_large SPLAT tests
were extended to provide better test coverage of various object
sizes and alignments.  Caches are created with random parameters
and their basic functionality is verified by allocating several
slabs worth of objects.

Signed-off-by: Brian Behlendorf <[email protected]>

2 files changed, 229 insertions, 140 deletions

diff --git a/module/spl/spl-kmem-cache.c b/module/spl/spl-kmem-cache.c
index 809ac5cc5..5160aa1cf 100644
--- a/module/spl/spl-kmem-cache.c
+++ b/module/spl/spl-kmem-cache.c
@@ -109,7 +109,7 @@ module_param(spl_kmem_cache_obj_per_slab_min, uint, 0644);
 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab_min,
 	"Minimal number of objects per slab");
 
-unsigned int spl_kmem_cache_max_size = 32;
+unsigned int spl_kmem_cache_max_size = SPL_KMEM_CACHE_MAX_SIZE;
 module_param(spl_kmem_cache_max_size, uint, 0644);
 MODULE_PARM_DESC(spl_kmem_cache_max_size, "Maximum size of slab in MB");
 
@@ -128,7 +128,13 @@ module_param(spl_kmem_cache_slab_limit, uint, 0644);
 MODULE_PARM_DESC(spl_kmem_cache_slab_limit,
 	"Objects less than N bytes use the Linux slab");
 
-unsigned int spl_kmem_cache_kmem_limit = (PAGE_SIZE / 4);
+/*
+ * This value defaults to a threshold designed to avoid allocations which
+ * have been deemed costly by the kernel.
+ */
+unsigned int spl_kmem_cache_kmem_limit =
+    ((1 << (PAGE_ALLOC_COSTLY_ORDER - 1)) * PAGE_SIZE) /
+    SPL_KMEM_CACHE_OBJ_PER_SLAB;
 module_param(spl_kmem_cache_kmem_limit, uint, 0644);
 MODULE_PARM_DESC(spl_kmem_cache_kmem_limit,
 	"Objects less than N bytes use the kmalloc");
@@ -181,12 +187,12 @@ kv_alloc(spl_kmem_cache_t *skc, int size, int flags)
 	gfp_t lflags = kmem_flags_convert(flags);
 	void *ptr;
 
-	ASSERT(ISP2(size));
-
-	if (skc->skc_flags & KMC_KMEM)
+	if (skc->skc_flags & KMC_KMEM) {
+		ASSERT(ISP2(size));
 		ptr = (void *)__get_free_pages(lflags, get_order(size));
-	else
+	} else {
 		ptr = spl_vmalloc(size, lflags | __GFP_HIGHMEM, PAGE_KERNEL);
+	}
 
 	/* Resulting allocated memory will be page aligned */
 	ASSERT(IS_P2ALIGNED(ptr, PAGE_SIZE));
@@ -198,7 +204,6 @@ static void
 kv_free(spl_kmem_cache_t *skc, void *ptr, int size)
 {
 	ASSERT(IS_P2ALIGNED(ptr, PAGE_SIZE));
-	ASSERT(ISP2(size));
 
 	/*
 	 * The Linux direct reclaim path uses this out of band value to
@@ -210,10 +215,12 @@ kv_free(spl_kmem_cache_t *skc, void *ptr, int size)
 	if (current->reclaim_state)
 		current->reclaim_state->reclaimed_slab += size >> PAGE_SHIFT;
 
-	if (skc->skc_flags & KMC_KMEM)
+	if (skc->skc_flags & KMC_KMEM) {
+		ASSERT(ISP2(size));
 		free_pages((unsigned long)ptr, get_order(size));
-	else
+	} else {
 		vfree(ptr);
+	}
 }
 
 /*
@@ -668,40 +675,48 @@ spl_cache_age(void *data)
 static int
 spl_slab_size(spl_kmem_cache_t *skc, uint32_t *objs, uint32_t *size)
 {
-	uint32_t sks_size, obj_size, max_size;
+	uint32_t sks_size, obj_size, max_size, tgt_size, tgt_objs;
 
 	if (skc->skc_flags & KMC_OFFSLAB) {
-		*objs = spl_kmem_cache_obj_per_slab;
-		*size = P2ROUNDUP(sizeof (spl_kmem_slab_t), PAGE_SIZE);
-		return (0);
+		tgt_objs = spl_kmem_cache_obj_per_slab;
+		tgt_size = P2ROUNDUP(sizeof (spl_kmem_slab_t), PAGE_SIZE);
+
+		if ((skc->skc_flags & KMC_KMEM) &&
+		    (spl_obj_size(skc) > (SPL_MAX_ORDER_NR_PAGES * PAGE_SIZE)))
+			return (-ENOSPC);
 	} else {
 		sks_size = spl_sks_size(skc);
 		obj_size = spl_obj_size(skc);
-
-		if (skc->skc_flags & KMC_KMEM)
-			max_size = ((uint32_t)1 << (MAX_ORDER-3)) * PAGE_SIZE;
-		else
-			max_size = (spl_kmem_cache_max_size * 1024 * 1024);
-
-		/* Power of two sized slab */
-		for (*size = PAGE_SIZE; *size <= max_size; *size *= 2) {
-			*objs = (*size - sks_size) / obj_size;
-			if (*objs >= spl_kmem_cache_obj_per_slab)
-				return (0);
-		}
+		max_size = (spl_kmem_cache_max_size * 1024 * 1024);
+		tgt_size = (spl_kmem_cache_obj_per_slab * obj_size + sks_size);
 
 		/*
-		 * Unable to satisfy target objects per slab, fall back to
-		 * allocating a maximally sized slab and assuming it can
-		 * contain the minimum objects count use it.  If not fail.
+		 * KMC_KMEM slabs are allocated by __get_free_pages() which
+		 * rounds up to the nearest order.  Knowing this the size
+		 * should be rounded up to the next power of two with a hard
+		 * maximum defined by the maximum allowed allocation order.
 		 */
-		*size = max_size;
-		*objs = (*size - sks_size) / obj_size;
-		if (*objs >= (spl_kmem_cache_obj_per_slab_min))
-			return (0);
+		if (skc->skc_flags & KMC_KMEM) {
+			max_size = SPL_MAX_ORDER_NR_PAGES * PAGE_SIZE;
+			tgt_size = MIN(max_size,
+			    PAGE_SIZE * (1 << MAX(get_order(tgt_size) - 1, 1)));
+		}
+
+		if (tgt_size <= max_size) {
+			tgt_objs = (tgt_size - sks_size) / obj_size;
+		} else {
+			tgt_objs = (max_size - sks_size) / obj_size;
+			tgt_size = (tgt_objs * obj_size) + sks_size;
+		}
 	}
 
-	return (-ENOSPC);
+	if (tgt_objs == 0)
+		return (-ENOSPC);
+
+	*objs = tgt_objs;
+	*size = tgt_size;
+
+	return (0);
 }
 
 /*
@@ -960,6 +975,11 @@ spl_kmem_cache_create(char *name, size_t size, size_t align,
 		if (rc)
 			goto out;
 	} else {
+		if (size > (SPL_MAX_KMEM_ORDER_NR_PAGES * PAGE_SIZE)) {
+			rc = EINVAL;
+			goto out;
+		}
+
 		skc->skc_linux_cache = kmem_cache_create(
 		    skc->skc_name, size, align, 0, NULL);
 		if (skc->skc_linux_cache == NULL) {
@@ -1406,8 +1426,11 @@ restart:
 		skm->skm_age = jiffies;
 	} else {
 		obj = spl_cache_refill(skc, skm, flags);
-		if (obj == NULL)
+		if ((obj == NULL) && !(flags & KM_NOSLEEP))
 			goto restart;
+
+		local_irq_enable();
+		goto ret;
 	}
 
 	local_irq_enable();
@@ -1427,7 +1450,6 @@ ret:
 
 	return (obj);
 }
-
 EXPORT_SYMBOL(spl_kmem_cache_alloc);
 
 /*
diff --git a/module/splat/splat-kmem.c b/module/splat/splat-kmem.c
index 81f748bb6..cd0000bae 100644
--- a/module/splat/splat-kmem.c
+++ b/module/splat/splat-kmem.c
@@ -27,6 +27,7 @@
 #include <sys/kmem.h>
 #include <sys/kmem_cache.h>
 #include <sys/vmem.h>
+#include <sys/random.h>
 #include <sys/thread.h>
 #include <sys/vmsystm.h>
 #include "splat-internal.h"
@@ -583,87 +584,124 @@ out:
 
 static int
 splat_kmem_cache_test(struct file *file, void *arg, char *name,
-		      int size, int align, int flags)
+    int size, int align, int flags)
 {
-	kmem_cache_priv_t *kcp;
-	kmem_cache_data_t *kcd = NULL;
-	int rc = 0, max;
+	kmem_cache_priv_t *kcp = NULL;
+	kmem_cache_data_t **kcd = NULL;
+	int i, rc = 0, objs = 0;
+
+	splat_vprint(file, name,
+	    "Testing size=%d, align=%d, flags=0x%04x\n",
+	    size, align, flags);
 
 	kcp = splat_kmem_cache_test_kcp_alloc(file, name, size, align, 0);
 	if (!kcp) {
 		splat_vprint(file, name, "Unable to create '%s'\n", "kcp");
-		return -ENOMEM;
+		return (-ENOMEM);
 	}
 
-	kcp->kcp_cache =
-		kmem_cache_create(SPLAT_KMEM_CACHE_NAME,
-				  kcp->kcp_size, kcp->kcp_align,
-				  splat_kmem_cache_test_constructor,
-				  splat_kmem_cache_test_destructor,
-				  NULL, kcp, NULL, flags);
-	if (!kcp->kcp_cache) {
-		splat_vprint(file, name,
-			     "Unable to create '%s'\n",
-			     SPLAT_KMEM_CACHE_NAME);
+	kcp->kcp_cache = kmem_cache_create(SPLAT_KMEM_CACHE_NAME,
+	    kcp->kcp_size, kcp->kcp_align,
+	    splat_kmem_cache_test_constructor,
+	    splat_kmem_cache_test_destructor,
+	    NULL, kcp, NULL, flags);
+	if (kcp->kcp_cache == NULL) {
+		splat_vprint(file, name, "Unable to create "
+		    "name='%s', size=%d, align=%d, flags=0x%x\n",
+		    SPLAT_KMEM_CACHE_NAME, size, align, flags);
 		rc = -ENOMEM;
 		goto out_free;
 	}
 
-	kcd = kmem_cache_alloc(kcp->kcp_cache, KM_SLEEP);
-	if (!kcd) {
-		splat_vprint(file, name,
-			     "Unable to allocate from '%s'\n",
-			     SPLAT_KMEM_CACHE_NAME);
-		rc = -EINVAL;
+	/*
+	 * Allocate several slabs worth of objects to verify functionality.
+	 * However, on 32-bit systems with limited address space constrain
+	 * it to a single slab for the purposes of this test.
+	 */
+#ifdef _LP64
+	objs = SPL_KMEM_CACHE_OBJ_PER_SLAB * 4;
+#else
+	objs = 1;
+#endif
+	kcd = kmem_zalloc(sizeof (kmem_cache_data_t *) * objs, KM_SLEEP);
+	if (kcd == NULL) {
+		splat_vprint(file, name, "Unable to allocate pointers "
+		    "for %d objects\n", objs);
+		rc = -ENOMEM;
 		goto out_free;
 	}
 
-	if (!kcd->kcd_flag) {
-		splat_vprint(file, name,
-			     "Failed to run contructor for '%s'\n",
-			     SPLAT_KMEM_CACHE_NAME);
-		rc = -EINVAL;
-		goto out_free;
-	}
+	for (i = 0; i < objs; i++) {
+		kcd[i] = kmem_cache_alloc(kcp->kcp_cache, KM_SLEEP);
+		if (kcd[i] == NULL) {
+			splat_vprint(file, name, "Unable to allocate "
+			    "from '%s'\n", SPLAT_KMEM_CACHE_NAME);
+			rc = -EINVAL;
+			goto out_free;
+		}
 
-	if (kcd->kcd_magic != kcp->kcp_magic) {
-		splat_vprint(file, name,
-			     "Failed to pass private data to constructor "
-			     "for '%s'\n", SPLAT_KMEM_CACHE_NAME);
-		rc = -EINVAL;
-		goto out_free;
+		if (!kcd[i]->kcd_flag) {
+			splat_vprint(file, name, "Failed to run constructor "
+			    "for '%s'\n", SPLAT_KMEM_CACHE_NAME);
+			rc = -EINVAL;
+			goto out_free;
+		}
+
+		if (kcd[i]->kcd_magic != kcp->kcp_magic) {
+			splat_vprint(file, name,
+			    "Failed to pass private data to constructor "
+			    "for '%s'\n", SPLAT_KMEM_CACHE_NAME);
+			rc = -EINVAL;
+			goto out_free;
+		}
 	}
 
-	max = kcp->kcp_count;
-	kmem_cache_free(kcp->kcp_cache, kcd);
+	for (i = 0; i < objs; i++) {
+		kmem_cache_free(kcp->kcp_cache, kcd[i]);
+
+		/* Destructors are run for every kmem_cache_free() */
+		if (kcd[i]->kcd_flag) {
+			splat_vprint(file, name,
+			    "Failed to run destructor for '%s'\n",
+			    SPLAT_KMEM_CACHE_NAME);
+			rc = -EINVAL;
+			goto out_free;
+		}
+	}
 
-	/* Destroy the entire cache which will force destructors to
-	 * run and we can verify one was called for every object */
-	kmem_cache_destroy(kcp->kcp_cache);
 	if (kcp->kcp_count) {
 		splat_vprint(file, name,
-			     "Failed to run destructor on all slab objects "
-			     "for '%s'\n", SPLAT_KMEM_CACHE_NAME);
+		    "Failed to run destructor on all slab objects for '%s'\n",
+		    SPLAT_KMEM_CACHE_NAME);
 		rc = -EINVAL;
 	}
 
+	kmem_free(kcd, sizeof (kmem_cache_data_t *) * objs);
+	kmem_cache_destroy(kcp->kcp_cache);
+
 	splat_kmem_cache_test_kcp_free(kcp);
 	splat_vprint(file, name,
-		     "Successfully ran ctors/dtors for %d elements in '%s'\n",
-		     max, SPLAT_KMEM_CACHE_NAME);
+	    "Success ran alloc'd/free'd %d objects of size %d\n",
+	    objs, size);
 
-	return rc;
+	return (rc);
 
 out_free:
-	if (kcd)
-		kmem_cache_free(kcp->kcp_cache, kcd);
+	if (kcd) {
+		for (i = 0; i < objs; i++) {
+			if (kcd[i] != NULL)
+				kmem_cache_free(kcp->kcp_cache, kcd[i]);
+		}
+
+		kmem_free(kcd, sizeof (kmem_cache_data_t *) * objs);
+	}
 
 	if (kcp->kcp_cache)
 		kmem_cache_destroy(kcp->kcp_cache);
 
 	splat_kmem_cache_test_kcp_free(kcp);
 
-	return rc;
+	return (rc);
 }
 
 static int
@@ -757,35 +795,49 @@ static int
 splat_kmem_test5(struct file *file, void *arg)
 {
 	char *name = SPLAT_KMEM_TEST5_NAME;
-	int rc;
-
-	/* On slab (default + kmem + vmem) */
-	rc = splat_kmem_cache_test(file, arg, name, 128, 0, 0);
-	if (rc)
-		return rc;
+	int i, rc = 0;
 
-	rc = splat_kmem_cache_test(file, arg, name, 128, 0, KMC_KMEM);
-	if (rc)
-		return rc;
+	/* Randomly pick small object sizes and alignments. */
+	for (i = 0; i < 100; i++) {
+		int size, align, flags = 0;
+		uint32_t rnd;
+
+		/* Evenly distribute tests over all value cache types */
+		get_random_bytes((void *)&rnd, sizeof (uint32_t));
+		switch (rnd & 0x03) {
+		default:
+		case 0x00:
+			flags = 0;
+			break;
+		case 0x01:
+			flags = KMC_KMEM;
+			break;
+		case 0x02:
+			flags = KMC_VMEM;
+			break;
+		case 0x03:
+			flags = KMC_SLAB;
+			break;
+		}
 
-	rc = splat_kmem_cache_test(file, arg, name, 128, 0, KMC_VMEM);
-	if (rc)
-		return rc;
+		/* The following flags are set with a 1/10 chance */
+		flags |= ((((rnd >> 8) % 10) == 0) ? KMC_OFFSLAB : 0);
+		flags |= ((((rnd >> 16) % 10) == 0) ? KMC_NOEMERGENCY : 0);
 
-	/* Off slab (default + kmem + vmem) */
-	rc = splat_kmem_cache_test(file, arg, name, 128, 0, KMC_OFFSLAB);
-	if (rc)
-		return rc;
+		/* 32b - PAGE_SIZE */
+		get_random_bytes((void *)&rnd, sizeof (uint32_t));
+		size = MAX(rnd % (PAGE_SIZE + 1), 32);
 
-	rc = splat_kmem_cache_test(file, arg, name, 128, 0,
-	    KMC_KMEM | KMC_OFFSLAB);
-	if (rc)
-		return rc;
+		/* 2^N where (3 <= N <= PAGE_SHIFT) */
+		get_random_bytes((void *)&rnd, sizeof (uint32_t));
+		align = (1 << MAX(3, rnd % (PAGE_SHIFT + 1)));
 
-	rc = splat_kmem_cache_test(file, arg, name, 128, 0,
-	    KMC_VMEM | KMC_OFFSLAB);
+		rc = splat_kmem_cache_test(file, arg, name, size, align, flags);
+		if (rc)
+			return (rc);
+	}
 
-	return rc;
+	return (rc);
 }
 
 /*
@@ -795,44 +847,53 @@ static int
 splat_kmem_test6(struct file *file, void *arg)
 {
 	char *name = SPLAT_KMEM_TEST6_NAME;
-	int rc;
-
-	/* On slab (default + kmem + vmem) */
-	rc = splat_kmem_cache_test(file, arg, name, 256*1024, 0, 0);
-	if (rc)
-		return rc;
-
-	rc = splat_kmem_cache_test(file, arg, name, 64*1024, 0, KMC_KMEM);
-	if (rc)
-		return rc;
-
-	rc = splat_kmem_cache_test(file, arg, name, 1024*1024, 0, KMC_VMEM);
-	if (rc)
-		return rc;
-
-	rc = splat_kmem_cache_test(file, arg, name, 16*1024*1024, 0, KMC_VMEM);
-	if (rc)
-		return rc;
+	int i, max_size, rc = 0;
+
+	/* Randomly pick large object sizes and alignments. */
+	for (i = 0; i < 100; i++) {
+		int size, align, flags = 0;
+		uint32_t rnd;
+
+		/* Evenly distribute tests over all value cache types */
+		get_random_bytes((void *)&rnd, sizeof (uint32_t));
+		switch (rnd & 0x03) {
+		default:
+		case 0x00:
+			flags = 0;
+			max_size = (SPL_KMEM_CACHE_MAX_SIZE * 1024 * 1024) / 2;
+			break;
+		case 0x01:
+			flags = KMC_KMEM;
+			max_size = (SPL_MAX_ORDER_NR_PAGES - 2) * PAGE_SIZE;
+			break;
+		case 0x02:
+			flags = KMC_VMEM;
+			max_size = (SPL_KMEM_CACHE_MAX_SIZE * 1024 * 1024) / 2;
+			break;
+		case 0x03:
+			flags = KMC_SLAB;
+			max_size = SPL_MAX_KMEM_ORDER_NR_PAGES * PAGE_SIZE;
+			break;
+		}
 
-	/* Off slab (default + kmem + vmem) */
-	rc = splat_kmem_cache_test(file, arg, name, 256*1024, 0, KMC_OFFSLAB);
-	if (rc)
-		return rc;
+		/* The following flags are set with a 1/10 chance */
+		flags |= ((((rnd >> 8) % 10) == 0) ? KMC_OFFSLAB : 0);
+		flags |= ((((rnd >> 16) % 10) == 0) ? KMC_NOEMERGENCY : 0);
 
-	rc = splat_kmem_cache_test(file, arg, name, 64*1024, 0,
-	    KMC_KMEM | KMC_OFFSLAB);
-	if (rc)
-		return rc;
+		/* PAGE_SIZE - max_size */
+		get_random_bytes((void *)&rnd, sizeof (uint32_t));
+		size = MAX(rnd % (max_size + 1), PAGE_SIZE),
 
-	rc = splat_kmem_cache_test(file, arg, name, 1024*1024, 0,
-	    KMC_VMEM | KMC_OFFSLAB);
-	if (rc)
-		return rc;
+		/* 2^N where (3 <= N <= PAGE_SHIFT) */
+		get_random_bytes((void *)&rnd, sizeof (uint32_t));
+		align = (1 << MAX(3, rnd % (PAGE_SHIFT + 1)));
 
-	rc = splat_kmem_cache_test(file, arg, name, 16*1024*1024, 0,
-	    KMC_VMEM | KMC_OFFSLAB);
+		rc = splat_kmem_cache_test(file, arg, name, size, align, flags);
+		if (rc)
+			return (rc);
+	}
 
-	return rc;
+	return (rc);
 }
 
 /*
@@ -842,14 +903,20 @@ static int
 splat_kmem_test7(struct file *file, void *arg)
 {
 	char *name = SPLAT_KMEM_TEST7_NAME;
+	int max_size = (SPL_KMEM_CACHE_MAX_SIZE * 1024 * 1024) / 2;
 	int i, rc;
 
 	for (i = SPL_KMEM_CACHE_ALIGN; i <= PAGE_SIZE; i *= 2) {
-		rc = splat_kmem_cache_test(file, arg, name, 157, i, 0);
+		uint32_t size;
+
+		get_random_bytes((void *)&size, sizeof (uint32_t));
+		size = MAX(size % (max_size + 1), 32);
+
+		rc = splat_kmem_cache_test(file, arg, name, size, i, 0);
 		if (rc)
 			return rc;
 
-		rc = splat_kmem_cache_test(file, arg, name, 157, i,
+		rc = splat_kmem_cache_test(file, arg, name, size, i,
 		    KMC_OFFSLAB);
 		if (rc)
 			return rc;