Re: [PATCH v2] mm/dmapool.c: avoid duplicate memset within dma_pool_alloc

From: Liu Song
Date: Tue Aug 16 2022 - 22:03:19 EST


On 2022-08-16 13:39, Marek Szyprowski wrote:
Hi,

On 18.07.2022 08:28, Liu Song wrote:
From: Liu Song <liusong@xxxxxxxxxxxxxxxxx>

In "dma_alloc_from_dev_coherent" and "dma_direct_alloc",
the allocated memory is explicitly set to 0.

A helper function "use_dev_coherent_memory" is introduced here to
determine whether the memory is allocated by "dma_alloc_from_dev_coherent".

And use "get_dma_ops" to determine whether the memory is allocated by
"dma_direct_alloc".

After this modification, memory allocated using "dma_pool_zalloc" can avoid
duplicate memset.

Signed-off-by: Liu Song <liusong@xxxxxxxxxxxxxxxxx>

This patch landed linux next-20220816. Unfortunately it causes serious
issues on ARM 32bit systems. I've observed it on ARM 32bit Samsung
Exynos 5422 based Odroid XU4 board with USB r8152 driver. After applying
this patch and loading r8152 driver I only the following endless
messages in the log:

xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs queued?
xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs queued?
xhci-hcd xhci-hcd.9.auto: WARN Event TRB for slot 1 ep 0 with no TDs queued?

It looks that there are drivers which rely on the fact that the dma
coherent buffers are always zeroed.

It's not even that, the change here is just obviously broken, since it ends up entirely ignoring want_init_on_alloc() for devices using dma-direct. Sure, the memory backing a dma_page is zeroed *once* by its initial dma-coherent allocation, but who says we're not not reallocating pool entries from an existing dma_page?

I'm not convinced it's worth trying to special-case this at all, since we can only do it reliably for the first pool entry allocated from a new dma_page, and that will only happen as the pool initially grows to a suitable size for its working set, after which no further new pages are likely to be allocated for the lifetime of the pool. Even if there is a case to be made for doing so, it would need to be based on the flow through dma_pool_alloc() itself, not some nonsense heuristic on the device.

Hi,

First of all, I am very sorry that there are missing branches that have not been considered fully,

but there is a possibility that the memset to 0 will be repeated whether re-allocation from the

dma pool, and this patch needs to be fixed.


Thanks



Andrew, please drop this patch.

Thanks,
Robin.

---
   include/linux/dma-map-ops.h | 5 +++++
   mm/dmapool.c                | 5 ++++-
   2 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/include/linux/dma-map-ops.h b/include/linux/dma-map-ops.h
index 0d5b06b..c29948d 100644
--- a/include/linux/dma-map-ops.h
+++ b/include/linux/dma-map-ops.h
@@ -171,6 +171,10 @@ int dma_alloc_from_dev_coherent(struct device *dev, ssize_t size,
   int dma_release_from_dev_coherent(struct device *dev, int order, void *vaddr);
   int dma_mmap_from_dev_coherent(struct device *dev, struct vm_area_struct *vma,
           void *cpu_addr, size_t size, int *ret);
+static inline bool use_dev_coherent_memory(struct device *dev)
+{
+    return dev->dma_mem ? true : false;
+}
   #else
   static inline int dma_declare_coherent_memory(struct device *dev,
           phys_addr_t phys_addr, dma_addr_t device_addr, size_t size)
@@ -180,6 +184,7 @@ static inline int dma_declare_coherent_memory(struct device *dev,
   #define dma_alloc_from_dev_coherent(dev, size, handle, ret) (0)
   #define dma_release_from_dev_coherent(dev, order, vaddr) (0)
   #define dma_mmap_from_dev_coherent(dev, vma, vaddr, order, ret) (0)
+#define use_dev_coherent_memory(dev) (0)
   #endif /* CONFIG_DMA_DECLARE_COHERENT */
      #ifdef CONFIG_DMA_GLOBAL_POOL
diff --git a/mm/dmapool.c b/mm/dmapool.c
index a7eb5d0..6e03530 100644
--- a/mm/dmapool.c
+++ b/mm/dmapool.c
@@ -21,6 +21,7 @@
      #include <linux/device.h>
   #include <linux/dma-mapping.h>
+#include <linux/dma-map-ops.h>
   #include <linux/dmapool.h>
   #include <linux/kernel.h>
   #include <linux/list.h>
@@ -372,7 +373,9 @@ void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
   #endif
       spin_unlock_irqrestore(&pool->lock, flags);
   -    if (want_init_on_alloc(mem_flags))
+    if (want_init_on_alloc(mem_flags) &&
+        !use_dev_coherent_memory(pool->dev) &&
+        get_dma_ops(pool->dev))
           memset(retval, 0, pool->size);
          return retval;

Best regards