* [dpdk-dev] [PATCH v2] contigmem: zero all pages during mmap
@ 2016-08-16 22:46 Jim Harris
2016-08-22 20:58 ` Thomas Monjalon
0 siblings, 1 reply; 2+ messages in thread
From: Jim Harris @ 2016-08-16 22:46 UTC (permalink / raw)
To: dev
On Linux, all huge pages are zeroed by the kernel before
first access by the DPDK application. But on FreeBSD,
the contigmem driver would only zero the contiguous
memory regions during initial driver load.
DPDK commit b78c91751 eliminated the explicit memset()
operation for rte_zmalloc(), which was OK on Linux
because the kernel zeroes the pages during app start,
but this broke FreeBSD. So this patch explicitly
zeroes the pages before they are mmap'd, to ensure
equivalent behavior to Linux
Fixes: b78c9175118f ("mem: do not zero out memory on zmalloc")
Reported-by: Daniel Verkamp <daniel.verkamp@intel.com>
Tested-by: Daniel Verkamp <daniel.verkamp@intel.com>
Acked-by: Sergio Gonzalez Monroy <sergio.gonzalez.monroy@intel.com>
Signed-off-by: Jim Harris <james.r.harris@intel.com>
---
lib/librte_eal/bsdapp/contigmem/contigmem.c | 8 ++++++--
1 file changed, 6 insertions(+), 2 deletions(-)
diff --git a/lib/librte_eal/bsdapp/contigmem/contigmem.c b/lib/librte_eal/bsdapp/contigmem/contigmem.c
index c6ca3b9..da971de 100644
--- a/lib/librte_eal/bsdapp/contigmem/contigmem.c
+++ b/lib/librte_eal/bsdapp/contigmem/contigmem.c
@@ -216,15 +216,19 @@ static int
contigmem_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t size,
struct vm_object **obj, int nprot)
{
+ uint64_t buffer_index;
+
/*
* The buffer index is encoded in the offset. Divide the offset by
* PAGE_SIZE to get the index of the buffer requested by the user
* app.
*/
- if ((*offset/PAGE_SIZE) >= contigmem_num_buffers)
+ buffer_index = *offset / PAGE_SIZE;
+ if (buffer_index >= contigmem_num_buffers)
return EINVAL;
- *offset = (vm_ooffset_t)vtophys(contigmem_buffers[*offset/PAGE_SIZE]);
+ memset(contigmem_buffers[buffer_index], 0, contigmem_buffer_size);
+ *offset = (vm_ooffset_t)vtophys(contigmem_buffers[buffer_index]);
*obj = vm_pager_allocate(OBJT_DEVICE, cdev, size, nprot, *offset,
curthread->td_ucred);
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [dpdk-dev] [PATCH v2] contigmem: zero all pages during mmap
2016-08-16 22:46 [dpdk-dev] [PATCH v2] contigmem: zero all pages during mmap Jim Harris
@ 2016-08-22 20:58 ` Thomas Monjalon
0 siblings, 0 replies; 2+ messages in thread
From: Thomas Monjalon @ 2016-08-22 20:58 UTC (permalink / raw)
To: Jim Harris; +Cc: dev, stable
2016-08-16 15:46, Jim Harris:
> On Linux, all huge pages are zeroed by the kernel before
> first access by the DPDK application. But on FreeBSD,
> the contigmem driver would only zero the contiguous
> memory regions during initial driver load.
>
> DPDK commit b78c91751 eliminated the explicit memset()
> operation for rte_zmalloc(), which was OK on Linux
> because the kernel zeroes the pages during app start,
> but this broke FreeBSD. So this patch explicitly
> zeroes the pages before they are mmap'd, to ensure
> equivalent behavior to Linux
Added precision that the bug is seen "when restarting app".
> Fixes: b78c9175118f ("mem: do not zero out memory on zmalloc")
>
> Reported-by: Daniel Verkamp <daniel.verkamp@intel.com>
> Tested-by: Daniel Verkamp <daniel.verkamp@intel.com>
> Acked-by: Sergio Gonzalez Monroy <sergio.gonzalez.monroy@intel.com>
> Signed-off-by: Jim Harris <james.r.harris@intel.com>
Applied, thanks
CC stable@, please pick this important fix in 16.07.
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2016-08-22 20:58 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-08-16 22:46 [dpdk-dev] [PATCH v2] contigmem: zero all pages during mmap Jim Harris
2016-08-22 20:58 ` Thomas Monjalon
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).