[AMD Official Use Only - General]
> -----Original Message-----
> From: Anoob Joseph <ano...@marvell.com>
> Sent: Wednesday, August 16, 2023 12:57 PM
> To: Varghese, Vipin <vipin.vargh...@amd.com>
> Cc: Yigit, Ferruh <ferruh.yi...@amd.com>; cheng1.ji...@intel.com;
> sta...@dpdk.org; tho...@monjalon.net; dev@dpdk.org; Jerin Jacob
> Kollanukkaran <jer...@marvell.com>
> Subject: RE: [EXT] [PATCH] app/dma-perf: fix physical address seg-fault
>
> Caution: This message originated from an External Source. Use proper caution
> when opening attachments, clicking links, or responding.
>
>
> Hi Vipin,
>
> Thanks for the update. Please see inline.
>
> Thanks,
> Anoob
>
> > -----Original Message-----
> > From: Vipin Varghese <vipin.vargh...@amd.com>
> > Sent: Wednesday, August 16, 2023 12:48 PM
> > To: tho...@monjalon.net; dev@dpdk.org
> > Cc: ferruh.yi...@amd.com; cheng1.ji...@intel.com; sta...@dpdk.org
> > Subject: [EXT] [PATCH] app/dma-perf: fix physical address seg-fault
> >
> > External Email
> >
> > ----------------------------------------------------------------------
> > do_cpu_mem_copy uses DPDK API rte_mbuf_data_iova to return the start
> > of the virtual address for both src and dst.
> > But in case of iova mode set as PA, this results in seg-fault.
> > This is because rte_memcpy VA address and not PA.
> >
> > this fix checks the iova mode and invokes rte_memcpy with the right
> > arguments.
> >
> > Bugzilla ID: 1269
> > Fixes: 623dc9364dc6 ("app/dma-perf: introduce DMA performance test")
> > Cc: cheng1.ji...@intel.com
> >
> > Cc: sta...@dpdk.org
> >
> > Signed-off-by: Vipin Varghese <vipin.vargh...@amd.com>
> > ---
> >
> > tested for both va and pa
> >
> > CMD:
> > PA: dpdk-test-dma-perf --iova-mode=pa -- --config test.ini
> > VA: dpdk-test-dma-perf --iova-mode=va -- --config test.ini
> > DC: dpdk-test-dma-perf --iova-mode=dc -- --config test.ini
> >
> > Log: fails for dc mode `EAL: invalid parameters for --iova-mode`
> >
> > test.ini:
> > ```
> > [case1]
> > type=CPU_MEM_COPY
> > mem_size=10
> > buf_size=64,8192,2,MUL
> > src_numa_node=0
> > dst_numa_node=0
> > cache_flush=0
> > test_seconds=2
> > lcore = 7
> > eal_args=--in-memory --no-pci
> > ```
> > ---
> > app/test-dma-perf/benchmark.c | 39 +++++++++++++++++++++++++++----
> > ----
> > 1 file changed, 30 insertions(+), 9 deletions(-)
> >
> > diff --git a/app/test-dma-perf/benchmark.c b/app/test-dma-
> > perf/benchmark.c index 0601e0d171..5573acc9f9 100644
> > --- a/app/test-dma-perf/benchmark.c
> > +++ b/app/test-dma-perf/benchmark.c
> > @@ -279,6 +279,10 @@ do_cpu_mem_copy(void *p)
> > struct rte_mbuf **srcs = para->srcs;
> > struct rte_mbuf **dsts = para->dsts;
> > uint32_t i;
> > + bool isAddrPaMode = false;
> > +
> > + if (rte_eal_iova_mode() == RTE_IOVA_PA)
> > + isAddrPaMode = true;
> >
> > worker_info->stop_flag = false;
> > worker_info->ready_flag = true;
> > @@ -286,16 +290,33 @@ do_cpu_mem_copy(void *p)
> > while (!worker_info->start_flag)
> > ;
> >
> > - while (1) {
> > - for (i = 0; i < nr_buf; i++) {
> > - /* copy buffer form src to dst */
> > - rte_memcpy((void
> > *)(uintptr_t)rte_mbuf_data_iova(dsts[i]),
> > - (void
> > *)(uintptr_t)rte_mbuf_data_iova(srcs[i]),
> > - (size_t)buf_size);
> > - worker_info->total_cpl++;
> > + if (true == isAddrPaMode) {
> > + while (1) {
> > + for (i = 0; i < nr_buf; i++) {
> > + void *src = rte_pktmbuf_mtod(dsts[i],
> > + void
> > *);
> > + void *dst = rte_pktmbuf_mtod(srcs[i],
> > + void
> > *);
> > +
> > + /* copy buffer form src to dst */
> > + rte_memcpy(dst,
> > + src,
> > + (size_t)buf_size);
> > + worker_info->total_cpl++;
> > + }
> > + if (worker_info->stop_flag)
> > + break;
> > + }
> > + } else {
> > + while (1) {
> > + for (i = 0; i < nr_buf; i++) {
> > + /* copy buffer form src to dst */
> > + rte_memcpy((void
> > *)(uintptr_t)rte_mbuf_data_iova(dsts[i]),
> > + (void
> > *)(uintptr_t)rte_mbuf_data_iova(srcs[i]),
> > + (size_t)buf_size);
>
> [Anoob] Even in case of VA as IOVA, rte_pktmbuf_mtod() should work, right? I
> was imagining a simple replacement of ' rte_mbuf_data_iova' with
> 'rte_pktmbuf_mtod' as the fix.
Ferruh and myself believe the rte_mbuf_data_iova should have been working for
both PA and VA. But internally does not.
I think what you are recommending is for both PA and VA just use `
rte_pktmbuf_mtod` for both. Thus we can avoid the check itself.
>
> > + worker_info->total_cpl++;
> > + }
> > + if (worker_info->stop_flag)
> > + break;
> > }
> > - if (worker_info->stop_flag)
> > - break;
> > }
> >
> > return 0;
> > --
> > 2.34.1
