Skip to contentSkip to conversation

regmap: Allocate buffers with GFP_ATOMIC when fast_io == true

PATCH Superseded Stephen Boyd 2 participants 4 messages Sep 10, 2015 – Sep 11, 2015

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 regmap: Allocate buffers with GFP_ATOMIC when fast_io == true

If a regmap is using fast_io, allocate buffers in the write APIs with GFP_ATOMIC instead of GFP_KERNEL. Otherwise we may schedule while atomic.

---
 drivers/base/regmap/internal.h |  1 +
 drivers/base/regmap/regmap.c   | 16 +++++++++++-----
 2 files changed, 12 insertions(+), 5 deletions(-)
drivers/base/regmap/internal.h +1 -0
index 873ddf91c9d3..3470ec146f12 100644
@@ -59,6 +59,7 @@ struct regmap {
 	regmap_lock lock;
 	regmap_unlock unlock;
 	void *lock_arg; /* This is passed to lock/unlock functions */
+	gfp_t alloc_flags;
 
 	struct device *dev; /* Device we do I/O on */
 	void *work_buf;     /* Scratch buffer used to format I/O */
drivers/base/regmap/regmap.c +11 -5
index 29128c6af445..0ad22f4d40b1 100644
@@ -563,6 +563,12 @@ struct regmap *regmap_init(struct device *dev,
 		}
 		map->lock_arg = map;
 	}
+
+	if ((bus && bus->fast_io) || config->fast_io)
+		map->alloc_flags = GFP_ATOMIC;
+	else
+		map->alloc_flags = GFP_KERNEL;
+
 	map->format.reg_bytes = DIV_ROUND_UP(config->reg_bits, 8);
 	map->format.pad_bytes = config->pad_bits / 8;
 	map->format.val_bytes = DIV_ROUND_UP(config->val_bits, 8);
@@ -1296,7 +1302,7 @@ int _regmap_raw_write(struct regmap *map, unsigned int reg,
 				return -ENOMEM;
 
 			async->work_buf = kzalloc(map->format.buf_size,
-						  GFP_KERNEL | GFP_DMA);
+						  map->alloc_flags | GFP_DMA);
 			if (!async->work_buf) {
 				kfree(async);
 				return -ENOMEM;
@@ -1358,7 +1364,7 @@ int _regmap_raw_write(struct regmap *map, unsigned int reg,
 	/* If that didn't work fall back on linearising by hand. */
 	if (ret == -ENOTSUPP) {
 		len = map->format.reg_bytes + map->format.pad_bytes + val_len;
-		buf = kzalloc(len, GFP_KERNEL);
+		buf = kzalloc(len, map->alloc_flags);
 		if (!buf)
 			return -ENOMEM;
 
@@ -1729,7 +1735,7 @@ out:
 		if (!val_count)
 			return -EINVAL;
 
-		wval = kmemdup(val, val_count * val_bytes, GFP_KERNEL);
+		wval = kmemdup(val, val_count * val_bytes, map->alloc_flags);
 		if (!wval) {
 			dev_err(map->dev, "Error in memory allocation\n");
 			return -ENOMEM;
@@ -1771,7 +1777,7 @@ static int _regmap_raw_multi_reg_write(struct regmap *map,
 	if (!len)
 		return -EINVAL;
 
-	buf = kzalloc(len, GFP_KERNEL);
+	buf = kzalloc(len, map->alloc_flags);
 	if (!buf)
 		return -ENOMEM;
 
@@ -1909,7 +1915,7 @@ static int _regmap_multi_reg_write(struct regmap *map,
 		if (range) {
 			size_t len = sizeof(struct reg_sequence)*num_regs;
 			struct reg_sequence *base = kmemdup(regs, len,
-							   GFP_KERNEL);
+							   map->alloc_flags);
 			if (!base)
 				return -ENOMEM;
 			ret = _regmap_range_multi_paged_reg_write(map, base,
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
maintainer
#1 lore raw
Quoted 3 lines · Stephen Boyd

If a regmap is using fast_io, allocate buffers in the write APIs with GFP_ATOMIC instead of GFP_KERNEL. Otherwise we may schedule while atomic.

Why is this needed? If something needs fast I/O it probably doesn't want to be going down any of the code paths that result in us doing allocations. I'd expect either no cache, a flat cache or setting up defaults at initialisation time.

#2 lore raw
Quoted 9 lines · Mark Brown
Quoted 3 lines · Stephen Boyd

If a regmap is using fast_io, allocate buffers in the write APIs with GFP_ATOMIC instead of GFP_KERNEL. Otherwise we may schedule while atomic.

Why is this needed? If something needs fast I/O it probably doesn't want to be going down any of the code paths that result in us doing allocations. I'd expect either no cache, a flat cache or setting up defaults at initialisation time.

We tripped over this with regmap_bulk_write() users on the SPMI bus. How about going down the same paths as !map->can_multi_write and map->use_single_rw if fast_io == true? Something like this untested patch.

drivers/base/regmap/regmap.c +2 -2
index 29128c6af445..4f2d75527c7f 100644
@@ -573,8 +573,8 @@ struct regmap *regmap_init(struct device *dev,
 		map->reg_stride = config->reg_stride;
 	else
 		map->reg_stride = 1;
-	map->use_single_rw = config->use_single_rw;
-	map->can_multi_write = config->can_multi_write;
+	map->use_single_rw = config->use_single_rw || config->fast_io;
+	map->can_multi_write = config->can_multi_write && !config->fast_io;
 	map->dev = dev;
 	map->bus = bus;
 	map->bus_context = bus_context;
Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
maintainer
#3 lore raw
Quoted 11 lines · Stephen Boyd
Quoted 4 lines · Mark Brown

Why is this needed? If something needs fast I/O it probably doesn't want to be going down any of the code paths that result in us doing allocations. I'd expect either no cache, a flat cache or setting up defaults at initialisation time.

We tripped over this with regmap_bulk_write() users on the SPMI bus. How about going down the same paths as !map->can_multi_write and map->use_single_rw if fast_io == true? Something like this untested patch.

Hrm, OK. Bulk writes are a reasonable use case here but we should be able to do better... if we had a cache we could bulk write directly from the cache but I don't think the SPMI devices tend to have caches so that won't help.

drivers/base/regmap/regmap.c struct regmap *regmap_init(struct device *dev, · line 577

Quoted 4 lines
drivers/base/regmap/regmap.c +2 -2
-	map->use_single_rw = config->use_single_rw;
-	map->can_multi_write = config->can_multi_write;
+	map->use_single_rw = config->use_single_rw || config->fast_io;
+	map->can_multi_write = config->can_multi_write && !config->fast_io;

The trouble with this is that you end up doing register at a time writes which isn't ideal for performance. Unless we can figure out a way to reliably allocate a bigger buffer ahead of time which I can't think of a way to do reliably without just putting a random number in the config (that we then end up either hitting or making too big) I think this is actually a reasonable use case.

Can you redo your patch to touch only the allocations in your data paths (I'm thinking particularly of the async changes as being unneeded, now I look again you didn't touch the cache code anyway) and put a comment in there about the alloc flags being used in fast path cases please?