of: Rework and simplify phandle cache to use a fixed size

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

---
 drivers/of/base.c       | 133 ++++++++--------------------------------
 drivers/of/dynamic.c    |   2 +-
 drivers/of/of_private.h |   4 +-
 drivers/of/overlay.c    |  10 ---
 4 files changed, 28 insertions(+), 121 deletions(-)

drivers/of/base.c +0 -5

-       if (phandle_cache) {
-               if (phandle_cache[masked_handle] &&
-                   handle == phandle_cache[masked_handle]->phandle)
-                       np = phandle_cache[masked_handle];
-               if (np && of_node_check_flag(np, OF_DETACHED)) {

Quoted 5 lines

drivers/of/base.c +0 -5

-       if (phandle_cache) {
-               if (phandle_cache[masked_handle] &&
-                   handle == phandle_cache[masked_handle]->phandle)
-                       np = phandle_cache[masked_handle];
-               if (np && of_node_check_flag(np, OF_DETACHED)) {

drivers/of/base.c struct device_node *of_find_node_by_phandle(phandle handle) · line 1177

Quoted 10 lines

drivers/of/base.c +5 -5

-                       WARN_ON(1); /* did not uncache np on node removal */
-                       of_node_put(np);
-                       phandle_cache[masked_handle] = NULL;
-                       np = NULL;
-               }
+       if (phandle_cache[handle_hash] &&
+           handle == phandle_cache[handle_hash]->phandle)
+               np = phandle_cache[handle_hash];
+       if (np && of_node_check_flag(np, OF_DETACHED)) {
+               WARN_ON(1); /* did not uncache np on node removal */

BTW, I don't think this check is even valid. If we failed to detach and remove the node from the cache, then we could be accessing np after freeing it.

Rob

Quoted 16 lines · Rob Herring

Quoted 5 lines

drivers/of/base.c +0 -5

-       if (phandle_cache) {
-               if (phandle_cache[masked_handle] &&
-                   handle == phandle_cache[masked_handle]->phandle)
-                       np = phandle_cache[masked_handle];
-               if (np && of_node_check_flag(np, OF_DETACHED)) {

drivers/of/base.c struct device_node *of_find_node_by_phandle(phandle handle) · line 1177

Quoted 10 lines

drivers/of/base.c +5 -5

-                       WARN_ON(1); /* did not uncache np on node removal */
-                       of_node_put(np);
-                       phandle_cache[masked_handle] = NULL;
-                       np = NULL;
-               }
+       if (phandle_cache[handle_hash] &&
+           handle == phandle_cache[handle_hash]->phandle)
+               np = phandle_cache[handle_hash];
+       if (np && of_node_check_flag(np, OF_DETACHED)) {
+               WARN_ON(1); /* did not uncache np on node removal */

Quoted 3 lines

BTW, I don't think this check is even valid. If we failed to detach and remove the node from the cache, then we could be accessing np after freeing it.

this is kmalloc()ed memory which is always valid. If the memory is already re-used then

        handle == phandle_cache[handle_hash]->phandle

will fail (the check, not the memory access itself).

If the check
remains valid then you can hope for the OF_DETACHED flag to trigger the warning.

Quoted 22 lines · Sebastian Andrzej Siewior

Quoted 19 lines · Rob Herring

Quoted 5 lines

drivers/of/base.c +0 -5

-       if (phandle_cache) {
-               if (phandle_cache[masked_handle] &&
-                   handle == phandle_cache[masked_handle]->phandle)
-                       np = phandle_cache[masked_handle];
-               if (np && of_node_check_flag(np, OF_DETACHED)) {

drivers/of/base.c struct device_node *of_find_node_by_phandle(phandle handle) · line 1177

Quoted 10 lines

drivers/of/base.c +5 -5

-                       WARN_ON(1); /* did not uncache np on node removal */
-                       of_node_put(np);
-                       phandle_cache[masked_handle] = NULL;
-                       np = NULL;
-               }
+       if (phandle_cache[handle_hash] &&
+           handle == phandle_cache[handle_hash]->phandle)
+               np = phandle_cache[handle_hash];
+       if (np && of_node_check_flag(np, OF_DETACHED)) {
+               WARN_ON(1); /* did not uncache np on node removal */

BTW, I don't think this check is even valid. If we failed to detach and remove the node from the cache, then we could be accessing np after freeing it.

Quoted 5 lines

this is kmalloc()ed memory which is always valid. If the memory is already re-used then

        handle == phandle_cache[handle_hash]->phandle

will fail (the check, not the memory access itself).

There's a 1 in 2^32 chance it won't.

Quoted 3 lines

If the check
remains valid then you can hope for the OF_DETACHED flag to trigger the warning.

Keyword is hope.

To look at it another way. Do we need this check? It is in the "fast path". There's a single location where we set OF_DETACHED and the cache entry is removed at the same time. Also, if we do free the node's memory, it also checks for OF_DETACHED. Previously, a free wouldn't happen because we incremented the ref count on nodes in the cache.

Rob

Quoted 16 lines · Rob Herring

Quoted 15 lines

drivers/of/base.c +5 -10

-       if (phandle_cache) {
-               if (phandle_cache[masked_handle] &&
-                   handle == phandle_cache[masked_handle]->phandle)
-                       np = phandle_cache[masked_handle];
-               if (np && of_node_check_flag(np, OF_DETACHED)) {
-                       WARN_ON(1); /* did not uncache np on node removal */
-                       of_node_put(np);
-                       phandle_cache[masked_handle] = NULL;
-                       np = NULL;
-               }
+       if (phandle_cache[handle_hash] &&
+           handle == phandle_cache[handle_hash]->phandle)
+               np = phandle_cache[handle_hash];
+       if (np && of_node_check_flag(np, OF_DETACHED)) {
+               WARN_ON(1); /* did not uncache np on node removal */

Quoted 3 lines

BTW, I don't think this check is even valid. If we failed to detach and remove the node from the cache, then we could be accessing np after freeing it.

this is kmalloc()ed memory which is always valid. If the memory is

already re-used then

	handle == phandle_cache[handle_hash]->phandle

will fail (the check, not the memory access itself). If the check remains valid then you can hope for the OF_DETACHED flag to trigger the warning.

Quoted 1 line

Rob

Sebastian

Quoted 60 lines · Rob Herring

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

---
 drivers/of/base.c       | 133 ++++++++--------------------------------
 drivers/of/dynamic.c    |   2 +-
 drivers/of/of_private.h |   4 +-
 drivers/of/overlay.c    |  10 ---
 4 files changed, 28 insertions(+), 121 deletions(-)

[...]

Quoted 5 lines

drivers/of/base.c +0 -5

-       if (phandle_cache) {
-               if (phandle_cache[masked_handle] &&
-                   handle == phandle_cache[masked_handle]->phandle)
-                       np = phandle_cache[masked_handle];
-               if (np && of_node_check_flag(np, OF_DETACHED)) {

drivers/of/base.c struct device_node *of_find_node_by_phandle(phandle handle) · line 1177

Quoted 10 lines

drivers/of/base.c +5 -5

-                       WARN_ON(1); /* did not uncache np on node removal */
-                       of_node_put(np);
-                       phandle_cache[masked_handle] = NULL;
-                       np = NULL;
-               }
+       if (phandle_cache[handle_hash] &&
+           handle == phandle_cache[handle_hash]->phandle)
+               np = phandle_cache[handle_hash];
+       if (np && of_node_check_flag(np, OF_DETACHED)) {
+               WARN_ON(1); /* did not uncache np on node removal */

BTW, I don't think this check is even valid. If we failed to detach and remove the node from the cache, then we could be accessing np after freeing it.

Rob

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

---
 drivers/of/base.c       | 133 ++++++++--------------------------------
 drivers/of/dynamic.c    |   2 +-
 drivers/of/of_private.h |   4 +-
 drivers/of/overlay.c    |  10 ---
 4 files changed, 28 insertions(+), 121 deletions(-)

drivers/of/base.c +26 -107

index db7fbc0c0893..f7da162f126d 100644
@@ -135,115 +135,38 @@ int __weak of_node_to_nid(struct device_node *np)
 }
 #endif
 
-/*
- * Assumptions behind phandle_cache implementation:
- *   - phandle property values are in a contiguous range of 1..n
- *
- * If the assumptions do not hold, then
- *   - the phandle lookup overhead reduction provided by the cache
- *     will likely be less
- */
+#define OF_PHANDLE_CACHE_BITS	7
+#define OF_PHANDLE_CACHE_SZ	BIT(OF_PHANDLE_CACHE_BITS)
 
-static struct device_node **phandle_cache;
-static u32 phandle_cache_mask;
+static struct device_node *phandle_cache[OF_PHANDLE_CACHE_SZ];
 
-/*
- * Caller must hold devtree_lock.
- */
-static void __of_free_phandle_cache(void)
+static u32 of_phandle_cache_hash(phandle handle)
 {
-	u32 cache_entries = phandle_cache_mask + 1;
-	u32 k;
-
-	if (!phandle_cache)
-		return;
-
-	for (k = 0; k < cache_entries; k++)
-		of_node_put(phandle_cache[k]);
-
-	kfree(phandle_cache);
-	phandle_cache = NULL;
+	return hash_32(handle, OF_PHANDLE_CACHE_BITS);
 }
 
-int of_free_phandle_cache(void)
-{
-	unsigned long flags;
-
-	raw_spin_lock_irqsave(&devtree_lock, flags);
-
-	__of_free_phandle_cache();
-
-	raw_spin_unlock_irqrestore(&devtree_lock, flags);
-
-	return 0;
-}
-#if !defined(CONFIG_MODULES)
-late_initcall_sync(of_free_phandle_cache);
-#endif
-
 /*
  * Caller must hold devtree_lock.
  */
-void __of_free_phandle_cache_entry(phandle handle)
+void __of_phandle_cache_inv_entry(phandle handle)
 {
-	phandle masked_handle;
+	u32 handle_hash;
 	struct device_node *np;
 
 	if (!handle)
 		return;
 
-	masked_handle = handle & phandle_cache_mask;
+	handle_hash = of_phandle_cache_hash(handle);
 
-	if (phandle_cache) {
-		np = phandle_cache[masked_handle];
-		if (np && handle == np->phandle) {
-			of_node_put(np);
-			phandle_cache[masked_handle] = NULL;
-		}
-	}
-}
-
-void of_populate_phandle_cache(void)
-{
-	unsigned long flags;
-	u32 cache_entries;
-	struct device_node *np;
-	u32 phandles = 0;
-
-	raw_spin_lock_irqsave(&devtree_lock, flags);
-
-	__of_free_phandle_cache();
-
-	for_each_of_allnodes(np)
-		if (np->phandle && np->phandle != OF_PHANDLE_ILLEGAL)
-			phandles++;
-
-	if (!phandles)
-		goto out;
-
-	cache_entries = roundup_pow_of_two(phandles);
-	phandle_cache_mask = cache_entries - 1;
-
-	phandle_cache = kcalloc(cache_entries, sizeof(*phandle_cache),
-				GFP_ATOMIC);
-	if (!phandle_cache)
-		goto out;
-
-	for_each_of_allnodes(np)
-		if (np->phandle && np->phandle != OF_PHANDLE_ILLEGAL) {
-			of_node_get(np);
-			phandle_cache[np->phandle & phandle_cache_mask] = np;
-		}
-
-out:
-	raw_spin_unlock_irqrestore(&devtree_lock, flags);
+	np = phandle_cache[handle_hash];
+	if (np && handle == np->phandle)
+		phandle_cache[handle_hash] = NULL;
 }
 
 void __init of_core_init(void)
 {
 	struct device_node *np;
 
-	of_populate_phandle_cache();
 
 	/* Create the kset, and register existing nodes */
 	mutex_lock(&of_mutex);

@@ -253,8 +176,11 @@ void __init of_core_init(void)
 		pr_err("failed to register existing nodes\n");
 		return;
 	}
-	for_each_of_allnodes(np)
+	for_each_of_allnodes(np) {
 		__of_attach_node_sysfs(np);
+		if (np->phandle && !phandle_cache[of_phandle_cache_hash(np->phandle)])
+			phandle_cache[of_phandle_cache_hash(np->phandle)] = np;
+	}
 	mutex_unlock(&of_mutex);
 
 	/* Symlink in /proc as required by userspace ABI */

@@ -1235,36 +1161,29 @@ struct device_node *of_find_node_by_phandle(phandle handle)
 {
 	struct device_node *np = NULL;
 	unsigned long flags;
-	phandle masked_handle;
+	u32 handle_hash;
 
 	if (!handle)
 		return NULL;
 
+	handle_hash = of_phandle_cache_hash(handle);
+
 	raw_spin_lock_irqsave(&devtree_lock, flags);
 
-	masked_handle = handle & phandle_cache_mask;
-
-	if (phandle_cache) {
-		if (phandle_cache[masked_handle] &&
-		    handle == phandle_cache[masked_handle]->phandle)
-			np = phandle_cache[masked_handle];
-		if (np && of_node_check_flag(np, OF_DETACHED)) {
-			WARN_ON(1); /* did not uncache np on node removal */
-			of_node_put(np);
-			phandle_cache[masked_handle] = NULL;
-			np = NULL;
-		}
+	if (phandle_cache[handle_hash] &&
+	    handle == phandle_cache[handle_hash]->phandle)
+		np = phandle_cache[handle_hash];
+	if (np && of_node_check_flag(np, OF_DETACHED)) {
+		WARN_ON(1); /* did not uncache np on node removal */
+		phandle_cache[handle_hash] = NULL;
+		np = NULL;
 	}
 
 	if (!np) {
 		for_each_of_allnodes(np)
 			if (np->phandle == handle &&
 			    !of_node_check_flag(np, OF_DETACHED)) {
-				if (phandle_cache) {
-					/* will put when removed from cache */
-					of_node_get(np);
-					phandle_cache[masked_handle] = np;
-				}
+				phandle_cache[handle_hash] = np;
 				break;
 			}
 	}

drivers/of/dynamic.c +1 -1

index 49b16f76d78e..08fd823edac9 100644
@@ -276,7 +276,7 @@ void __of_detach_node(struct device_node *np)
 	of_node_set_flag(np, OF_DETACHED);
 
 	/* race with of_find_node_by_phandle() prevented by devtree_lock */
-	__of_free_phandle_cache_entry(np->phandle);
+	__of_phandle_cache_inv_entry(np->phandle);
 }
 
 /**

drivers/of/of_private.h +1 -3

index 66294d29942a..582844c158ae 100644
@@ -85,14 +85,12 @@ int of_resolve_phandles(struct device_node *tree);
 #endif
 
 #if defined(CONFIG_OF_DYNAMIC)
-void __of_free_phandle_cache_entry(phandle handle);
+void __of_phandle_cache_inv_entry(phandle handle);
 #endif
 
 #if defined(CONFIG_OF_OVERLAY)
 void of_overlay_mutex_lock(void);
 void of_overlay_mutex_unlock(void);
-int of_free_phandle_cache(void);
-void of_populate_phandle_cache(void);
 #else
 static inline void of_overlay_mutex_lock(void) {};
 static inline void of_overlay_mutex_unlock(void) {};

drivers/of/overlay.c +0 -10

index 9617b7df7c4d..97fe92c1f1d2 100644
@@ -974,8 +974,6 @@ static int of_overlay_apply(const void *fdt, struct device_node *tree,
 		goto err_free_overlay_changeset;
 	}
 
-	of_populate_phandle_cache();
-
 	ret = __of_changeset_apply_notify(&ovcs->cset);
 	if (ret)
 		pr_err("overlay apply changeset entry notify error %d\n", ret);

@@ -1218,17 +1216,9 @@ int of_overlay_remove(int *ovcs_id)
 
 	list_del(&ovcs->ovcs_list);
 
-	/*
-	 * Disable phandle cache.  Avoids race condition that would arise
-	 * from removing cache entry when the associated node is deleted.
-	 */
-	of_free_phandle_cache();
-
 	ret_apply = 0;
 	ret = __of_changeset_revert_entries(&ovcs->cset, &ret_apply);
 
-	of_populate_phandle_cache();
-
 	if (ret) {
 		if (ret_apply)
 			devicetree_state_flags |= DTSF_REVERT_FAIL;

+Chintan

Hi Chintan,

Quoted 15 lines · Rob Herring

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Quoted 5 lines

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Can you check this patch on the system that you reported phandle lookup overhead issues for?

-Frank

Quoted 297 lines

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

---
 drivers/of/base.c       | 133 ++++++++--------------------------------
 drivers/of/dynamic.c    |   2 +-
 drivers/of/of_private.h |   4 +-
 drivers/of/overlay.c    |  10 ---
 4 files changed, 28 insertions(+), 121 deletions(-)

drivers/of/base.c +26 -107

index db7fbc0c0893..f7da162f126d 100644
@@ -135,115 +135,38 @@ int __weak of_node_to_nid(struct device_node *np)
 }
 #endif
 
-/*
- * Assumptions behind phandle_cache implementation:
- *   - phandle property values are in a contiguous range of 1..n
- *
- * If the assumptions do not hold, then
- *   - the phandle lookup overhead reduction provided by the cache
- *     will likely be less
- */
+#define OF_PHANDLE_CACHE_BITS	7
+#define OF_PHANDLE_CACHE_SZ	BIT(OF_PHANDLE_CACHE_BITS)
 
-static struct device_node **phandle_cache;
-static u32 phandle_cache_mask;
+static struct device_node *phandle_cache[OF_PHANDLE_CACHE_SZ];
 
-/*
- * Caller must hold devtree_lock.
- */
-static void __of_free_phandle_cache(void)
+static u32 of_phandle_cache_hash(phandle handle)
 {
-	u32 cache_entries = phandle_cache_mask + 1;
-	u32 k;
-
-	if (!phandle_cache)
-		return;
-
-	for (k = 0; k < cache_entries; k++)
-		of_node_put(phandle_cache[k]);
-
-	kfree(phandle_cache);
-	phandle_cache = NULL;
+	return hash_32(handle, OF_PHANDLE_CACHE_BITS);
 }
 
-int of_free_phandle_cache(void)
-{
-	unsigned long flags;
-
-	raw_spin_lock_irqsave(&devtree_lock, flags);
-
-	__of_free_phandle_cache();
-
-	raw_spin_unlock_irqrestore(&devtree_lock, flags);
-
-	return 0;
-}
-#if !defined(CONFIG_MODULES)
-late_initcall_sync(of_free_phandle_cache);
-#endif
-
 /*
  * Caller must hold devtree_lock.
  */
-void __of_free_phandle_cache_entry(phandle handle)
+void __of_phandle_cache_inv_entry(phandle handle)
 {
-	phandle masked_handle;
+	u32 handle_hash;
 	struct device_node *np;
 
 	if (!handle)
 		return;
 
-	masked_handle = handle & phandle_cache_mask;
+	handle_hash = of_phandle_cache_hash(handle);
 
-	if (phandle_cache) {
-		np = phandle_cache[masked_handle];
-		if (np && handle == np->phandle) {
-			of_node_put(np);
-			phandle_cache[masked_handle] = NULL;
-		}
-	}
-}
-
-void of_populate_phandle_cache(void)
-{
-	unsigned long flags;
-	u32 cache_entries;
-	struct device_node *np;
-	u32 phandles = 0;
-
-	raw_spin_lock_irqsave(&devtree_lock, flags);
-
-	__of_free_phandle_cache();
-
-	for_each_of_allnodes(np)
-		if (np->phandle && np->phandle != OF_PHANDLE_ILLEGAL)
-			phandles++;
-
-	if (!phandles)
-		goto out;
-
-	cache_entries = roundup_pow_of_two(phandles);
-	phandle_cache_mask = cache_entries - 1;
-
-	phandle_cache = kcalloc(cache_entries, sizeof(*phandle_cache),
-				GFP_ATOMIC);
-	if (!phandle_cache)
-		goto out;
-
-	for_each_of_allnodes(np)
-		if (np->phandle && np->phandle != OF_PHANDLE_ILLEGAL) {
-			of_node_get(np);
-			phandle_cache[np->phandle & phandle_cache_mask] = np;
-		}
-
-out:
-	raw_spin_unlock_irqrestore(&devtree_lock, flags);
+	np = phandle_cache[handle_hash];
+	if (np && handle == np->phandle)
+		phandle_cache[handle_hash] = NULL;
 }
 
 void __init of_core_init(void)
 {
 	struct device_node *np;
 
-	of_populate_phandle_cache();
 
 	/* Create the kset, and register existing nodes */
 	mutex_lock(&of_mutex);

@@ -253,8 +176,11 @@ void __init of_core_init(void)
 		pr_err("failed to register existing nodes\n");
 		return;
 	}
-	for_each_of_allnodes(np)
+	for_each_of_allnodes(np) {
 		__of_attach_node_sysfs(np);
+		if (np->phandle && !phandle_cache[of_phandle_cache_hash(np->phandle)])
+			phandle_cache[of_phandle_cache_hash(np->phandle)] = np;
+	}
 	mutex_unlock(&of_mutex);
 
 	/* Symlink in /proc as required by userspace ABI */

@@ -1235,36 +1161,29 @@ struct device_node *of_find_node_by_phandle(phandle handle)
 {
 	struct device_node *np = NULL;
 	unsigned long flags;
-	phandle masked_handle;
+	u32 handle_hash;
 
 	if (!handle)
 		return NULL;
 
+	handle_hash = of_phandle_cache_hash(handle);
+
 	raw_spin_lock_irqsave(&devtree_lock, flags);
 
-	masked_handle = handle & phandle_cache_mask;
-
-	if (phandle_cache) {
-		if (phandle_cache[masked_handle] &&
-		    handle == phandle_cache[masked_handle]->phandle)
-			np = phandle_cache[masked_handle];
-		if (np && of_node_check_flag(np, OF_DETACHED)) {
-			WARN_ON(1); /* did not uncache np on node removal */
-			of_node_put(np);
-			phandle_cache[masked_handle] = NULL;
-			np = NULL;
-		}
+	if (phandle_cache[handle_hash] &&
+	    handle == phandle_cache[handle_hash]->phandle)
+		np = phandle_cache[handle_hash];
+	if (np && of_node_check_flag(np, OF_DETACHED)) {
+		WARN_ON(1); /* did not uncache np on node removal */
+		phandle_cache[handle_hash] = NULL;
+		np = NULL;
 	}
 
 	if (!np) {
 		for_each_of_allnodes(np)
 			if (np->phandle == handle &&
 			    !of_node_check_flag(np, OF_DETACHED)) {
-				if (phandle_cache) {
-					/* will put when removed from cache */
-					of_node_get(np);
-					phandle_cache[masked_handle] = np;
-				}
+				phandle_cache[handle_hash] = np;
 				break;
 			}
 	}

drivers/of/dynamic.c +1 -1

index 49b16f76d78e..08fd823edac9 100644
@@ -276,7 +276,7 @@ void __of_detach_node(struct device_node *np)
 	of_node_set_flag(np, OF_DETACHED);
 
 	/* race with of_find_node_by_phandle() prevented by devtree_lock */
-	__of_free_phandle_cache_entry(np->phandle);
+	__of_phandle_cache_inv_entry(np->phandle);
 }
 
 /**

drivers/of/of_private.h +1 -3

index 66294d29942a..582844c158ae 100644
@@ -85,14 +85,12 @@ int of_resolve_phandles(struct device_node *tree);
 #endif
 
 #if defined(CONFIG_OF_DYNAMIC)
-void __of_free_phandle_cache_entry(phandle handle);
+void __of_phandle_cache_inv_entry(phandle handle);
 #endif
 
 #if defined(CONFIG_OF_OVERLAY)
 void of_overlay_mutex_lock(void);
 void of_overlay_mutex_unlock(void);
-int of_free_phandle_cache(void);
-void of_populate_phandle_cache(void);
 #else
 static inline void of_overlay_mutex_lock(void) {};
 static inline void of_overlay_mutex_unlock(void) {};

drivers/of/overlay.c +0 -10

index 9617b7df7c4d..97fe92c1f1d2 100644
@@ -974,8 +974,6 @@ static int of_overlay_apply(const void *fdt, struct device_node *tree,
 		goto err_free_overlay_changeset;
 	}
 
-	of_populate_phandle_cache();
-
 	ret = __of_changeset_apply_notify(&ovcs->cset);
 	if (ret)
 		pr_err("overlay apply changeset entry notify error %d\n", ret);

@@ -1218,17 +1216,9 @@ int of_overlay_remove(int *ovcs_id)
 
 	list_del(&ovcs->ovcs_list);
 
-	/*
-	 * Disable phandle cache.  Avoids race condition that would arise
-	 * from removing cache entry when the associated node is deleted.
-	 */
-	of_free_phandle_cache();
-
 	ret_apply = 0;
 	ret = __of_changeset_revert_entries(&ovcs->cset, &ret_apply);
 
-	of_populate_phandle_cache();
-
 	if (ret) {
 		if (ret_apply)
 			devicetree_state_flags |= DTSF_REVERT_FAIL;

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

---
 drivers/of/base.c       | 133 ++++++++--------------------------------
 drivers/of/dynamic.c    |   2 +-
 drivers/of/of_private.h |   4 +-
 drivers/of/overlay.c    |  10 ---
 4 files changed, 28 insertions(+), 121 deletions(-)

drivers/of/base.c +26 -107

index db7fbc0c0893..f7da162f126d 100644
@@ -135,115 +135,38 @@ int __weak of_node_to_nid(struct device_node *np)
 }
 #endif
 
-/*
- * Assumptions behind phandle_cache implementation:
- *   - phandle property values are in a contiguous range of 1..n
- *
- * If the assumptions do not hold, then
- *   - the phandle lookup overhead reduction provided by the cache
- *     will likely be less
- */
+#define OF_PHANDLE_CACHE_BITS	7
+#define OF_PHANDLE_CACHE_SZ	BIT(OF_PHANDLE_CACHE_BITS)
 
-static struct device_node **phandle_cache;
-static u32 phandle_cache_mask;
+static struct device_node *phandle_cache[OF_PHANDLE_CACHE_SZ];
 
-/*
- * Caller must hold devtree_lock.
- */
-static void __of_free_phandle_cache(void)
+static u32 of_phandle_cache_hash(phandle handle)
 {
-	u32 cache_entries = phandle_cache_mask + 1;
-	u32 k;
-
-	if (!phandle_cache)
-		return;
-
-	for (k = 0; k < cache_entries; k++)
-		of_node_put(phandle_cache[k]);
-
-	kfree(phandle_cache);
-	phandle_cache = NULL;
+	return hash_32(handle, OF_PHANDLE_CACHE_BITS);
 }
 
-int of_free_phandle_cache(void)
-{
-	unsigned long flags;
-
-	raw_spin_lock_irqsave(&devtree_lock, flags);
-
-	__of_free_phandle_cache();
-
-	raw_spin_unlock_irqrestore(&devtree_lock, flags);
-
-	return 0;
-}
-#if !defined(CONFIG_MODULES)
-late_initcall_sync(of_free_phandle_cache);
-#endif
-
 /*
  * Caller must hold devtree_lock.
  */
-void __of_free_phandle_cache_entry(phandle handle)
+void __of_phandle_cache_inv_entry(phandle handle)
 {
-	phandle masked_handle;
+	u32 handle_hash;
 	struct device_node *np;
 
 	if (!handle)
 		return;
 
-	masked_handle = handle & phandle_cache_mask;
+	handle_hash = of_phandle_cache_hash(handle);
 
-	if (phandle_cache) {
-		np = phandle_cache[masked_handle];
-		if (np && handle == np->phandle) {
-			of_node_put(np);
-			phandle_cache[masked_handle] = NULL;
-		}
-	}
-}
-
-void of_populate_phandle_cache(void)
-{
-	unsigned long flags;
-	u32 cache_entries;
-	struct device_node *np;
-	u32 phandles = 0;
-
-	raw_spin_lock_irqsave(&devtree_lock, flags);
-
-	__of_free_phandle_cache();
-
-	for_each_of_allnodes(np)
-		if (np->phandle && np->phandle != OF_PHANDLE_ILLEGAL)
-			phandles++;
-
-	if (!phandles)
-		goto out;
-
-	cache_entries = roundup_pow_of_two(phandles);
-	phandle_cache_mask = cache_entries - 1;
-
-	phandle_cache = kcalloc(cache_entries, sizeof(*phandle_cache),
-				GFP_ATOMIC);
-	if (!phandle_cache)
-		goto out;
-
-	for_each_of_allnodes(np)
-		if (np->phandle && np->phandle != OF_PHANDLE_ILLEGAL) {
-			of_node_get(np);
-			phandle_cache[np->phandle & phandle_cache_mask] = np;
-		}
-
-out:
-	raw_spin_unlock_irqrestore(&devtree_lock, flags);
+	np = phandle_cache[handle_hash];
+	if (np && handle == np->phandle)
+		phandle_cache[handle_hash] = NULL;
 }
 
 void __init of_core_init(void)
 {
 	struct device_node *np;
 
-	of_populate_phandle_cache();
 
 	/* Create the kset, and register existing nodes */
 	mutex_lock(&of_mutex);

@@ -253,8 +176,11 @@ void __init of_core_init(void)
 		pr_err("failed to register existing nodes\n");
 		return;
 	}
-	for_each_of_allnodes(np)
+	for_each_of_allnodes(np) {
 		__of_attach_node_sysfs(np);
+		if (np->phandle && !phandle_cache[of_phandle_cache_hash(np->phandle)])
+			phandle_cache[of_phandle_cache_hash(np->phandle)] = np;
+	}
 	mutex_unlock(&of_mutex);
 
 	/* Symlink in /proc as required by userspace ABI */

@@ -1235,36 +1161,29 @@ struct device_node *of_find_node_by_phandle(phandle handle)
 {
 	struct device_node *np = NULL;
 	unsigned long flags;
-	phandle masked_handle;
+	u32 handle_hash;
 
 	if (!handle)
 		return NULL;
 
+	handle_hash = of_phandle_cache_hash(handle);
+
 	raw_spin_lock_irqsave(&devtree_lock, flags);
 
-	masked_handle = handle & phandle_cache_mask;
-
-	if (phandle_cache) {
-		if (phandle_cache[masked_handle] &&
-		    handle == phandle_cache[masked_handle]->phandle)
-			np = phandle_cache[masked_handle];
-		if (np && of_node_check_flag(np, OF_DETACHED)) {
-			WARN_ON(1); /* did not uncache np on node removal */
-			of_node_put(np);
-			phandle_cache[masked_handle] = NULL;
-			np = NULL;
-		}
+	if (phandle_cache[handle_hash] &&
+	    handle == phandle_cache[handle_hash]->phandle)
+		np = phandle_cache[handle_hash];
+	if (np && of_node_check_flag(np, OF_DETACHED)) {
+		WARN_ON(1); /* did not uncache np on node removal */
+		phandle_cache[handle_hash] = NULL;
+		np = NULL;
 	}
 
 	if (!np) {
 		for_each_of_allnodes(np)
 			if (np->phandle == handle &&
 			    !of_node_check_flag(np, OF_DETACHED)) {
-				if (phandle_cache) {
-					/* will put when removed from cache */
-					of_node_get(np);
-					phandle_cache[masked_handle] = np;
-				}
+				phandle_cache[handle_hash] = np;
 				break;
 			}
 	}

drivers/of/dynamic.c +1 -1

index 49b16f76d78e..08fd823edac9 100644
@@ -276,7 +276,7 @@ void __of_detach_node(struct device_node *np)
 	of_node_set_flag(np, OF_DETACHED);
 
 	/* race with of_find_node_by_phandle() prevented by devtree_lock */
-	__of_free_phandle_cache_entry(np->phandle);
+	__of_phandle_cache_inv_entry(np->phandle);
 }
 
 /**

drivers/of/of_private.h +0 -0

index 66294d29942a..582844c158ae 100644

drivers/of/of_private.h +0 -2

 #if defined(CONFIG_OF_OVERLAY)
 void of_overlay_mutex_lock(void);
 void of_overlay_mutex_unlock(void);
-int of_free_phandle_cache(void);
-void of_populate_phandle_cache(void);
 #else
 static inline void of_overlay_mutex_lock(void) {};
 static inline void of_overlay_mutex_unlock(void) {};

drivers/of/overlay.c +0 -6

index 9617b7df7c4d..97fe92c1f1d2 100644
@@ -974,8 +974,6 @@ static int of_overlay_apply(const void *fdt, struct device_node *tree,
 		goto err_free_overlay_changeset;
 	}
 
-	of_populate_phandle_cache();
-
 	ret = __of_changeset_apply_notify(&ovcs->cset);
 	if (ret)
 		pr_err("overlay apply changeset entry notify error %d\n", ret);

@@ -1218,17 +1216,9 @@ int of_overlay_remove(int *ovcs_id)
 
 	list_del(&ovcs->ovcs_list);
 
-	/*
-	 * Disable phandle cache.  Avoids race condition that would arise
-	 * from removing cache entry when the associated node is deleted.
-	 */

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

Hi Rob,

Quoted 48 lines · Rob Herring

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Quoted 5 lines

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <mpe-Gsx/[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

With next-20200106 I have noticed a regression on Tegra210 where it appears that only one of the eMMC devices is being registered. Bisect is pointing to this patch and reverting on top of next fixes the problem. That is as far as I have got so far, so if you have any ideas, please let me know. Unfortunately, there do not appear to be any obvious errors from the bootlog.

Quoted 58 lines · Jon Hunter

Hi Rob,

Quoted 48 lines · Rob Herring

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Quoted 5 lines

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <mpe-Gsx/[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

Quoted 6 lines

With next-20200106 I have noticed a regression on Tegra210 where it appears that only one of the eMMC devices is being registered. Bisect is pointing to this patch and reverting on top of next fixes the problem. That is as far as I have got so far, so if you have any ideas, please let me know. Unfortunately, there do not appear to be any obvious errors from the bootlog.

I guess that's tegra210-p2371-2180.dts because none of the others have 2 SD hosts enabled. I don't see anything obvious though. Are you doing any runtime mods to the DT?

Quoted 66 lines · Rob Herring

Quoted 64 lines · Jon Hunter

Hi Rob,

Quoted 48 lines · Rob Herring

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Quoted 5 lines

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <mpe-Gsx/[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

With next-20200106 I have noticed a regression on Tegra210 where it appears that only one of the eMMC devices is being registered. Bisect is pointing to this patch and reverting on top of next fixes the problem. That is as far as I have got so far, so if you have any ideas, please let me know. Unfortunately, there do not appear to be any obvious errors from the bootlog.

Quoted 3 lines

I guess that's tegra210-p2371-2180.dts because none of the others have 2 SD hosts enabled. I don't see anything obvious though. Are you doing any runtime mods to the DT?

I have noticed that the bootloader is doing some runtime mods and so checking if this is the cause. I will let you know, but most likely, seeing as I cannot find anything wrong with this change itself.

+Karol

Quoted 69 lines · Jon Hunter

Quoted 66 lines · Rob Herring

Quoted 64 lines · Jon Hunter

Hi Rob,

Quoted 48 lines · Rob Herring

The phandle cache was added to speed up of_find_node_by_phandle() by avoiding walking the whole DT to find a matching phandle. The implementation has several shortcomings:

- The cache is designed to work on a linear set of phandle values. This is true for dtc generated DTs, but not for other cases such as Power.
- The cache isn't enabled until of_core_init() and a typical system may see hundreds of calls to of_find_node_by_phandle() before that point.
- The cache is freed and re-allocated when the number of phandles changes.
- It takes a raw spinlock around a memory allocation which breaks on

    RT.

Change the implementation to a fixed size and use hash_32() as the cache index. This greatly simplifies the implementation. It avoids the need for any re-alloc of the cache and taking a reference on nodes in the cache. We only have a single source of removing cache entries which is of_detach_node().

Using hash_32() removes any assumption on phandle values improving the hit rate for non-linear phandle values. The effect on linear values using hash_32() is about a 10% collision. The chances of thrashing on colliding values seems to be low.

To compare performance, I used a RK3399 board which is a pretty typical system. I found that just measuring boot time as done previously is noisy and may be impacted by other things. Also bringing up secondary cores causes some issues with measuring, so I booted with 'nr_cpus=1'. With no caching, calls to of_find_node_by_phandle() take about 20124 us for 1248 calls. There's an additional 288 calls before time keeping is up. Using the average time per hit/miss with the cache, we can calculate these calls to take 690 us (277 hit / 11 miss) with a 128 entry cache and 13319 us with no cache or an uninitialized cache.

Comparing the 3 implementations the time spent in
of_find_node_by_phandle() is:

no cache: 20124 us (+ 13319 us)
128 entry cache: 5134 us (+ 690 us)
current cache: 819 us (+ 13319 us)

We could move the allocation of the cache earlier to improve the current cache, but that just further complicates the situation as it needs to be after slab is up, so we can't do it when unflattening (which uses memblock).

Quoted 5 lines

Reported-by Sebastian Andrzej Siewior <[email protected]>

Cc Michael Ellerman <mpe-Gsx/[email protected]>

Cc Segher Boessenkool <[email protected]>

Cc Frank Rowand <[email protected]>

Signed-off-by Rob Herring <[email protected]>

With next-20200106 I have noticed a regression on Tegra210 where it appears that only one of the eMMC devices is being registered. Bisect is pointing to this patch and reverting on top of next fixes the problem. That is as far as I have got so far, so if you have any ideas, please let me know. Unfortunately, there do not appear to be any obvious errors from the bootlog.

Quoted 7 lines

Quoted 3 lines

I guess that's tegra210-p2371-2180.dts because none of the others have 2 SD hosts enabled. I don't see anything obvious though. Are you doing any runtime mods to the DT?

I have noticed that the bootloader is doing some runtime mods and so checking if this is the cause. I will let you know, but most likely, seeing as I cannot find anything wrong with this change itself.

Did you figure out the problem here? Karol sees a similar problem on Tegra210 with the gpu node regulator.

It looks like /external-memory-controller@7001b000 has a duplicate phandle. Comparing the dtb in the filesystem with what the kernel gets, that node is added by the bootloader. So the bootloader is definitely creating a broken dtb.