Skip to contentSkip to conversation

get rid of dma_max_pfn

PATCH In review Christoph Hellwig 2 participants 5 messages Apr 11, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Hi everyone,

I though I got rid of all non-highmem, non-ISA block layer bounce buffering a while ago, but I missed the MMC case. While I still plan to also kill off the highmem bouncing there I won't get to it this merge window, so for now I'd like to make some progress and move MMC to the DMA layer (swiotlb or arm dmabounce) bouncing for addressing limitations and kill off the dma_max_pfn helper.

#1 lore raw

PATCH 1/2 mmc: let the dma map ops handle bouncing

Just like we do for all other block drivers. Especially as the limit imposed at the moment might be way to pessimistic for iommus.

---
 drivers/mmc/core/queue.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
drivers/mmc/core/queue.c +2 -5
index 7c364a9c4eeb..eb9c0692062c 100644
@@ -354,18 +354,15 @@ static const struct blk_mq_ops mmc_mq_ops = {
 static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
 {
 	struct mmc_host *host = card->host;
-	u64 limit = BLK_BOUNCE_HIGH;
 	unsigned block_size = 512;
 
-	if (mmc_dev(host)->dma_mask && *mmc_dev(host)->dma_mask)
-		limit = (u64)dma_max_pfn(mmc_dev(host)) << PAGE_SHIFT;
-
 	blk_queue_flag_set(QUEUE_FLAG_NONROT, mq->queue);
 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
 	if (mmc_can_erase(card))
 		mmc_queue_setup_discard(mq->queue, card);
 
-	blk_queue_bounce_limit(mq->queue, limit);
+	if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask)
+		blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);
 	blk_queue_max_hw_sectors(mq->queue,
 		min(host->max_blk_count, host->max_req_size / 512));
 	blk_queue_max_segments(mq->queue, host->max_segs);
Signature · 1 line
2.20.1
#2 lore raw

Re: [PATCH 1/2] mmc: let the dma map ops handle bouncing

Hi Christoph,

Quoted 3 lines · Christoph Hellwig

Just like we do for all other block drivers. Especially as the limit imposed at the moment might be way to pessimistic for iommus.

I would appreciate some information in the changelog, as it's quite unclear of what this change really means.

Quoted 18 lines
---
 drivers/mmc/core/queue.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
drivers/mmc/core/queue.c +0 -2
index 7c364a9c4eeb..eb9c0692062c 100644
@@ -354,18 +354,15 @@ static const struct blk_mq_ops mmc_mq_ops = {
 static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
 {
        struct mmc_host *host = card->host;
-       u64 limit = BLK_BOUNCE_HIGH;
        unsigned block_size = 512;
 
-       if (mmc_dev(host)->dma_mask && *mmc_dev(host)->dma_mask)

drivers/mmc/core/queue.c static const struct blk_mq_ops mmc_mq_ops = { · line 365

Quoted 10 lines
drivers/mmc/core/queue.c +2 -3
-               limit = (u64)dma_max_pfn(mmc_dev(host)) << PAGE_SHIFT;
-
        blk_queue_flag_set(QUEUE_FLAG_NONROT, mq->queue);
        blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
        if (mmc_can_erase(card))
                mmc_queue_setup_discard(mq->queue, card);
 
-       blk_queue_bounce_limit(mq->queue, limit);
+       if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask)
+               blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);

So this means we are not going to set a bounce limit for the queue, in case we have a dma mask.

Why isn't that needed no more? Whats has changed?

Quoted 6 lines
drivers/mmc/core/queue.c +0 -1
        blk_queue_max_hw_sectors(mq->queue,
                min(host->max_blk_count, host->max_req_size / 512));
        blk_queue_max_segments(mq->queue, host->max_segs);
--

2.20.1

Kind regards
Uffe

#3 lore raw

Re: [PATCH 1/2] mmc: let the dma map ops handle bouncing

Quoted 4 lines · Ulf Hansson
Quoted 4 lines
        blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
        if (mmc_can_erase(card))
                mmc_queue_setup_discard(mq->queue, card);
Quoted 8 lines

drivers/mmc/core/queue.c static const struct blk_mq_ops mmc_mq_ops = { · line 365

Quoted 3 lines
drivers/mmc/core/queue.c +2 -1
-       blk_queue_bounce_limit(mq->queue, limit);
+       if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask)
+               blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);

So this means we are not going to set a bounce limit for the queue, in case we have a dma mask.

Why isn't that needed no more? Whats has changed?

On most architectures it was never needed, the major hold out was x86-32 with PAE. In general the dma_mask tells the DMA API layer what is supported, and if the physical addressing doesn't support that it has to use bounce buffering like swiotlb (or dmabounce on arm32). A couple month ago I finally fixes x86-32 to also properly set up swiotlb, and remove the block layerer bounce buffering that wasn't for highmem (which is about having a kernel mapping, not addressing), and ISA DMA (which is not handled like everything else, but we'll get there). But for some reason I missed mmc back then, so mmc right now is the only remaining user of address based block layer bouncing.

#4 lore raw

PATCH 2/2 dma-mapping: remove dma_max_pfn

In reply to Christoph Hellwig

These days the DMA mapping code must bounce buffer for any not supported address, and if they driver needs to optimize for natively supported ranged it should use dma_get_required_mask.

---
 arch/arm/include/asm/dma-mapping.h | 7 -------
 include/linux/dma-mapping.h        | 7 -------
 2 files changed, 14 deletions(-)
arch/arm/include/asm/dma-mapping.h +0 -7
index 03ba90ffc0f8..7e0486ad1318 100644
@@ -89,13 +89,6 @@ static inline dma_addr_t virt_to_dma(struct device *dev, void *addr)
 }
 #endif
 
-/* The ARM override for dma_max_pfn() */
-static inline unsigned long dma_max_pfn(struct device *dev)
-{
-	return dma_to_pfn(dev, *dev->dma_mask);
-}
-#define dma_max_pfn(dev) dma_max_pfn(dev)
-
 /* do not use this function in a driver */
 static inline bool is_device_dma_coherent(struct device *dev)
 {
include/linux/dma-mapping.h +0 -7
index 75e60be91e5f..6976d36c05de 100644
@@ -729,13 +729,6 @@ static inline int dma_set_seg_boundary(struct device *dev, unsigned long mask)
 	return -EIO;
 }
 
-#ifndef dma_max_pfn
-static inline unsigned long dma_max_pfn(struct device *dev)
-{
-	return (*dev->dma_mask >> PAGE_SHIFT) + dev->dma_pfn_offset;
-}
-#endif
-
 static inline int dma_get_cache_alignment(void)
 {
 #ifdef ARCH_DMA_MINALIGN
Signature · 1 line
2.20.1