With modern NIC, it is not unusual having about ~256,000 active dma
mappings. Hash size of 1024 buckets is too small.
Forcing full cache line per bucket does not seem useful,
especially now that we have a contention on free_entries_lock
for allocations and freeing of entries. Better using space
to fit more buckets.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Marek Szyprowski <[email protected]>
---
kernel/dma/debug.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index 4ad74f5987ea9e95f9bb5e2d1592254e367d24fb..35e2a853bff9c482d789ab331d79aaee07753a97 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -27,7 +27,7 @@
#include <asm/sections.h>
-#define HASH_SIZE 1024ULL
+#define HASH_SIZE 16384ULL
#define HASH_FN_SHIFT 13
#define HASH_FN_MASK (HASH_SIZE - 1)
@@ -87,7 +87,7 @@ typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
struct hash_bucket {
struct list_head list;
spinlock_t lock;
-} ____cacheline_aligned_in_smp;
+};
/* Hash list to save the allocated dma addresses */
static struct hash_bucket dma_entry_hash[HASH_SIZE];
--
2.24.0.rc0.303.g954a862665-goog
On Wed, Oct 30, 2019 at 8:13 PM Eric Dumazet <[email protected]> wrote:
> With modern NIC, it is not unusual having about ~256,000 active dma
> mappings. Hash size of 1024 buckets is too small.
>
> Forcing full cache line per bucket does not seem useful,
> especially now that we have a contention on free_entries_lock
> for allocations and freeing of entries. Better using space
> to fit more buckets.
>
> Signed-off-by: Eric Dumazet <[email protected]>
> Cc: Christoph Hellwig <[email protected]>
> Cc: Marek Szyprowski <[email protected]>
> ---
> kernel/dma/debug.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
> index 4ad74f5987ea9e95f9bb5e2d1592254e367d24fb..35e2a853bff9c482d789ab331d79aaee07753a97 100644
> --- a/kernel/dma/debug.c
> +++ b/kernel/dma/debug.c
> @@ -27,7 +27,7 @@
>
> #include <asm/sections.h>
>
> -#define HASH_SIZE 1024ULL
> +#define HASH_SIZE 16384ULL
> #define HASH_FN_SHIFT 13
> #define HASH_FN_MASK (HASH_SIZE - 1)
>
> @@ -87,7 +87,7 @@ typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
> struct hash_bucket {
> struct list_head list;
> spinlock_t lock;
> -} ____cacheline_aligned_in_smp;
> +};
>
> /* Hash list to save the allocated dma addresses */
> static struct hash_bucket dma_entry_hash[HASH_SIZE];
JFTR, this increases dma_entry_hash size by 327680 bytes, and pushes
a few more boards beyond their bootloader-imposed kernel size limits.
Disabling CONFIG_DMA_API_DEBUG fixes that.
Of course the real fix is to fix the bootloaders...
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
On Tue, Dec 10, 2019 at 6:55 AM Geert Uytterhoeven <[email protected]> wrote:
> JFTR, this increases dma_entry_hash size by 327680 bytes, and pushes
> a few more boards beyond their bootloader-imposed kernel size limits.
>
> Disabling CONFIG_DMA_API_DEBUG fixes that.
> Of course the real fix is to fix the bootloaders...
>
Hi Geert
Maybe we can make the hash size arch-dependent, or better dynamically
allocate this memory ?
Hi Eric,
On Tue, Dec 10, 2019 at 4:04 PM Eric Dumazet <[email protected]> wrote:
> On Tue, Dec 10, 2019 at 6:55 AM Geert Uytterhoeven <[email protected]> wrote:
> > JFTR, this increases dma_entry_hash size by 327680 bytes, and pushes
> > a few more boards beyond their bootloader-imposed kernel size limits.
> >
> > Disabling CONFIG_DMA_API_DEBUG fixes that.
> > Of course the real fix is to fix the bootloaders...
>
> Maybe we can make the hash size arch-dependent, or better dynamically
> allocate this memory ?
Dynamically allocating would be the better solution, IMHO.
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
On Tue, Dec 10, 2019 at 03:55:08PM +0100, Geert Uytterhoeven wrote:
> On Wed, Oct 30, 2019 at 8:13 PM Eric Dumazet <[email protected]> wrote:
> > With modern NIC, it is not unusual having about ~256,000 active dma
> > mappings. Hash size of 1024 buckets is too small.
> >
> > Forcing full cache line per bucket does not seem useful,
> > especially now that we have a contention on free_entries_lock
> > for allocations and freeing of entries. Better using space
> > to fit more buckets.
> >
> > Signed-off-by: Eric Dumazet <[email protected]>
> > Cc: Christoph Hellwig <[email protected]>
> > Cc: Marek Szyprowski <[email protected]>
> > ---
> > kernel/dma/debug.c | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
> > index 4ad74f5987ea9e95f9bb5e2d1592254e367d24fb..35e2a853bff9c482d789ab331d79aaee07753a97 100644
> > --- a/kernel/dma/debug.c
> > +++ b/kernel/dma/debug.c
> > @@ -27,7 +27,7 @@
> >
> > #include <asm/sections.h>
> >
> > -#define HASH_SIZE 1024ULL
> > +#define HASH_SIZE 16384ULL
> > #define HASH_FN_SHIFT 13
> > #define HASH_FN_MASK (HASH_SIZE - 1)
> >
> > @@ -87,7 +87,7 @@ typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
> > struct hash_bucket {
> > struct list_head list;
> > spinlock_t lock;
> > -} ____cacheline_aligned_in_smp;
> > +};
> >
> > /* Hash list to save the allocated dma addresses */
> > static struct hash_bucket dma_entry_hash[HASH_SIZE];
>
> JFTR, this increases dma_entry_hash size by 327680 bytes, and pushes
> a few more boards beyond their bootloader-imposed kernel size limits.
>
> Disabling CONFIG_DMA_API_DEBUG fixes that.
> Of course the real fix is to fix the bootloaders...
Can someone just send a patch to switch this to a dynamic allocation?
On Thu, Jan 30, 2020 at 8:27 AM Christoph Hellwig <[email protected]> wrote:
>
> On Tue, Dec 10, 2019 at 03:55:08PM +0100, Geert Uytterhoeven wrote:
>
> Can someone just send a patch to switch this to a dynamic allocation?
Yes I can do that, thanks for the reminder.