peer->outbuf is a virtual address which is get by ioremap, it can not
be converted to a physical address by virt_to_page and page_to_phys.
This conversion will result in DMA error, because the destination address
which is converted by page_to_phys is invalid.
We Save the physical address in perf_setup_peer_mw, it is MMIO address
of NTB BARx. Then fill the destination address of DMA descriptor with
this physical address to guarantee that the address of memory write
requests fall in memory window of NBT BARx.
Signed-off-by: Jiasen Lin <[email protected]>
---
drivers/ntb/test/ntb_perf.c | 28 +++++++++++++++++++++++-----
1 file changed, 23 insertions(+), 5 deletions(-)
diff --git a/drivers/ntb/test/ntb_perf.c b/drivers/ntb/test/ntb_perf.c
index e9b7c2d..1c2fd1a 100644
--- a/drivers/ntb/test/ntb_perf.c
+++ b/drivers/ntb/test/ntb_perf.c
@@ -149,6 +149,7 @@ struct perf_peer {
u64 outbuf_xlat;
resource_size_t outbuf_size;
void __iomem *outbuf;
+ phys_addr_t out_phys_addr;
/* Inbound MW params */
dma_addr_t inbuf_xlat;
@@ -776,7 +777,8 @@ static void perf_dma_copy_callback(void *data)
}
static int perf_copy_chunk(struct perf_thread *pthr,
- void __iomem *dst, void *src, size_t len)
+ void __iomem *dst, void *src, size_t len,
+ phys_addr_t dst_phys_addr)
{
struct dma_async_tx_descriptor *tx;
struct dmaengine_unmap_data *unmap;
@@ -807,8 +809,7 @@ static int perf_copy_chunk(struct perf_thread *pthr,
}
unmap->to_cnt = 1;
- unmap->addr[1] = dma_map_page(dma_dev, virt_to_page(dst),
- offset_in_page(dst), len, DMA_FROM_DEVICE);
+ unmap->addr[1] = dst_phys_addr;
if (dma_mapping_error(dma_dev, unmap->addr[1])) {
ret = -EIO;
goto err_free_resource;
@@ -901,6 +902,8 @@ static int perf_run_test(struct perf_thread *pthr)
u64 total_size, chunk_size;
void *flt_src;
int ret = 0;
+ phys_addr_t flt_phys_addr;
+ phys_addr_t bnd_phys_addr;
total_size = 1ULL << total_order;
chunk_size = 1ULL << chunk_order;
@@ -909,12 +912,15 @@ static int perf_run_test(struct perf_thread *pthr)
flt_src = pthr->src;
bnd_dst = peer->outbuf + peer->outbuf_size;
flt_dst = peer->outbuf;
+ bnd_phys_addr = peer->out_phys_addr + peer->outbuf_size;
+ flt_phys_addr = peer->out_phys_addr;
pthr->duration = ktime_get();
/* Copied field is cleared on test launch stage */
while (pthr->copied < total_size) {
- ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size);
+ ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size,
+ flt_phys_addr);
if (ret) {
dev_err(&perf->ntb->dev, "%d: Got error %d on test\n",
pthr->tidx, ret);
@@ -925,8 +931,15 @@ static int perf_run_test(struct perf_thread *pthr)
flt_dst += chunk_size;
flt_src += chunk_size;
- if (flt_dst >= bnd_dst || flt_dst < peer->outbuf) {
+ flt_phys_addr += chunk_size;
+
+ if (flt_dst >= bnd_dst ||
+ flt_dst < peer->outbuf ||
+ flt_phys_addr >= bnd_phys_addr ||
+ flt_phys_addr < peer->out_phys_addr) {
+
flt_dst = peer->outbuf;
+ flt_phys_addr = peer->out_phys_addr;
flt_src = pthr->src;
}
@@ -1195,6 +1208,9 @@ static ssize_t perf_dbgfs_read_info(struct file *filep, char __user *ubuf,
"\tOut buffer addr 0x%pK\n", peer->outbuf);
pos += scnprintf(buf + pos, buf_size - pos,
+ "\tOut buff phys addr %pa[p]\n", &peer->out_phys_addr);
+
+ pos += scnprintf(buf + pos, buf_size - pos,
"\tOut buffer size %pa\n", &peer->outbuf_size);
pos += scnprintf(buf + pos, buf_size - pos,
@@ -1388,6 +1404,8 @@ static int perf_setup_peer_mw(struct perf_peer *peer)
if (!peer->outbuf)
return -ENOMEM;
+ peer->out_phys_addr = phys_addr;
+
if (max_mw_size && peer->outbuf_size > max_mw_size) {
peer->outbuf_size = max_mw_size;
dev_warn(&peer->perf->ntb->dev,
--
2.7.4
On 2019-11-06 8:38 p.m., Jiasen Lin wrote:
> peer->outbuf is a virtual address which is get by ioremap, it can not
> be converted to a physical address by virt_to_page and page_to_phys.
> This conversion will result in DMA error, because the destination address
> which is converted by page_to_phys is invalid.
Hmm, yes, ntb_perf is obviously wrong. I never noticed this, how did
this ever work?
> We Save the physical address in perf_setup_peer_mw, it is MMIO address
> of NTB BARx. Then fill the destination address of DMA descriptor with
> this physical address to guarantee that the address of memory write
> requests fall in memory window of NBT BARx.
Using the physical address directly is also wrong and will not work in
the presence of an IOMMU. You should use dma_map_resource() for this.
See what was done in ntb_transport[1].
Thanks,
Logan
[1] c59666bb32b9 ("NTB: ntb_transport: Ensure the destination buffer is
mapped for TX DMA")
> Signed-off-by: Jiasen Lin <[email protected]>
> ---
> drivers/ntb/test/ntb_perf.c | 28 +++++++++++++++++++++++-----
> 1 file changed, 23 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/ntb/test/ntb_perf.c b/drivers/ntb/test/ntb_perf.c
> index e9b7c2d..1c2fd1a 100644
> --- a/drivers/ntb/test/ntb_perf.c
> +++ b/drivers/ntb/test/ntb_perf.c
> @@ -149,6 +149,7 @@ struct perf_peer {
> u64 outbuf_xlat;
> resource_size_t outbuf_size;
> void __iomem *outbuf;
> + phys_addr_t out_phys_addr;
>
> /* Inbound MW params */
> dma_addr_t inbuf_xlat;
> @@ -776,7 +777,8 @@ static void perf_dma_copy_callback(void *data)
> }
>
> static int perf_copy_chunk(struct perf_thread *pthr,
> - void __iomem *dst, void *src, size_t len)
> + void __iomem *dst, void *src, size_t len,
> + phys_addr_t dst_phys_addr)
> {
> struct dma_async_tx_descriptor *tx;
> struct dmaengine_unmap_data *unmap;
> @@ -807,8 +809,7 @@ static int perf_copy_chunk(struct perf_thread *pthr,
> }
> unmap->to_cnt = 1;
>
> - unmap->addr[1] = dma_map_page(dma_dev, virt_to_page(dst),
> - offset_in_page(dst), len, DMA_FROM_DEVICE);
> + unmap->addr[1] = dst_phys_addr;
> if (dma_mapping_error(dma_dev, unmap->addr[1])) {
> ret = -EIO;
> goto err_free_resource;
> @@ -901,6 +902,8 @@ static int perf_run_test(struct perf_thread *pthr)
> u64 total_size, chunk_size;
> void *flt_src;
> int ret = 0;
> + phys_addr_t flt_phys_addr;
> + phys_addr_t bnd_phys_addr;
>
> total_size = 1ULL << total_order;
> chunk_size = 1ULL << chunk_order;
> @@ -909,12 +912,15 @@ static int perf_run_test(struct perf_thread *pthr)
> flt_src = pthr->src;
> bnd_dst = peer->outbuf + peer->outbuf_size;
> flt_dst = peer->outbuf;
> + bnd_phys_addr = peer->out_phys_addr + peer->outbuf_size;
> + flt_phys_addr = peer->out_phys_addr;
>
> pthr->duration = ktime_get();
>
> /* Copied field is cleared on test launch stage */
> while (pthr->copied < total_size) {
> - ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size);
> + ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size,
> + flt_phys_addr);
> if (ret) {
> dev_err(&perf->ntb->dev, "%d: Got error %d on test\n",
> pthr->tidx, ret);
> @@ -925,8 +931,15 @@ static int perf_run_test(struct perf_thread *pthr)
>
> flt_dst += chunk_size;
> flt_src += chunk_size;
> - if (flt_dst >= bnd_dst || flt_dst < peer->outbuf) {
> + flt_phys_addr += chunk_size;
> +
> + if (flt_dst >= bnd_dst ||
> + flt_dst < peer->outbuf ||
> + flt_phys_addr >= bnd_phys_addr ||
> + flt_phys_addr < peer->out_phys_addr) {
> +
> flt_dst = peer->outbuf;
> + flt_phys_addr = peer->out_phys_addr;
> flt_src = pthr->src;
> }
>
> @@ -1195,6 +1208,9 @@ static ssize_t perf_dbgfs_read_info(struct file *filep, char __user *ubuf,
> "\tOut buffer addr 0x%pK\n", peer->outbuf);
>
> pos += scnprintf(buf + pos, buf_size - pos,
> + "\tOut buff phys addr %pa[p]\n", &peer->out_phys_addr);
> +
> + pos += scnprintf(buf + pos, buf_size - pos,
> "\tOut buffer size %pa\n", &peer->outbuf_size);
>
> pos += scnprintf(buf + pos, buf_size - pos,
> @@ -1388,6 +1404,8 @@ static int perf_setup_peer_mw(struct perf_peer *peer)
> if (!peer->outbuf)
> return -ENOMEM;
>
> + peer->out_phys_addr = phys_addr;
> +
> if (max_mw_size && peer->outbuf_size > max_mw_size) {
> peer->outbuf_size = max_mw_size;
> dev_warn(&peer->perf->ntb->dev,
>
On 2019/11/9 1:04, Logan Gunthorpe wrote:
>
>
> On 2019-11-06 8:38 p.m., Jiasen Lin wrote:
>> peer->outbuf is a virtual address which is get by ioremap, it can not
>> be converted to a physical address by virt_to_page and page_to_phys.
>> This conversion will result in DMA error, because the destination address
>> which is converted by page_to_phys is invalid.
>
> Hmm, yes, ntb_perf is obviously wrong. I never noticed this, how did
> this ever work?
>
The default value of use_dma which is used to enable DMA engine to
measure NTB performance is zero, in other words, DMA engine is not used
by default. Thus, olny memcpy_toio is called in perf_copy_chunk and not
trigger this bug.
If we install driver with specified dma-enabled flag like this:
insmod ntb_perf.ko use_dma=1, this issue will be triggered.
>> We Save the physical address in perf_setup_peer_mw, it is MMIO address
>> of NTB BARx. Then fill the destination address of DMA descriptor with
>> this physical address to guarantee that the address of memory write
>> requests fall in memory window of NBT BARx.
>
> Using the physical address directly is also wrong and will not work in
> the presence of an IOMMU. You should use dma_map_resource() for this.
> See what was done in ntb_transport[1].
>
Yes, my mistake. I will modify the code according to your suggestion and
test it on AMD and HYGON platforms with the IOMMU enabled. Maybe the
following patches are relied on, when IOMMU is enabled on AMD and HYGON
plartforms.
https://lore.kernel.org/patchwork/cover/1143155/
https://lore.kernel.org/patchwork/patch/1143156/
https://lore.kernel.org/patchwork/patch/1143157/
Thanks,
Jiasen Lin
> Thanks,
>
> Logan
>
> [1] c59666bb32b9 ("NTB: ntb_transport: Ensure the destination buffer is
> mapped for TX DMA")
>
>> Signed-off-by: Jiasen Lin <[email protected]>
>> ---
>> drivers/ntb/test/ntb_perf.c | 28 +++++++++++++++++++++++-----
>> 1 file changed, 23 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/ntb/test/ntb_perf.c b/drivers/ntb/test/ntb_perf.c
>> index e9b7c2d..1c2fd1a 100644
>> --- a/drivers/ntb/test/ntb_perf.c
>> +++ b/drivers/ntb/test/ntb_perf.c
>> @@ -149,6 +149,7 @@ struct perf_peer {
>> u64 outbuf_xlat;
>> resource_size_t outbuf_size;
>> void __iomem *outbuf;
>> + phys_addr_t out_phys_addr;
>>
>> /* Inbound MW params */
>> dma_addr_t inbuf_xlat;
>> @@ -776,7 +777,8 @@ static void perf_dma_copy_callback(void *data)
>> }
>>
>> static int perf_copy_chunk(struct perf_thread *pthr,
>> - void __iomem *dst, void *src, size_t len)
>> + void __iomem *dst, void *src, size_t len,
>> + phys_addr_t dst_phys_addr)
>> {
>> struct dma_async_tx_descriptor *tx;
>> struct dmaengine_unmap_data *unmap;
>> @@ -807,8 +809,7 @@ static int perf_copy_chunk(struct perf_thread *pthr,
>> }
>> unmap->to_cnt = 1;
>>
>> - unmap->addr[1] = dma_map_page(dma_dev, virt_to_page(dst),
>> - offset_in_page(dst), len, DMA_FROM_DEVICE);
>> + unmap->addr[1] = dst_phys_addr;
>> if (dma_mapping_error(dma_dev, unmap->addr[1])) {
>> ret = -EIO;
>> goto err_free_resource;
>> @@ -901,6 +902,8 @@ static int perf_run_test(struct perf_thread *pthr)
>> u64 total_size, chunk_size;
>> void *flt_src;
>> int ret = 0;
>> + phys_addr_t flt_phys_addr;
>> + phys_addr_t bnd_phys_addr;
>>
>> total_size = 1ULL << total_order;
>> chunk_size = 1ULL << chunk_order;
>> @@ -909,12 +912,15 @@ static int perf_run_test(struct perf_thread *pthr)
>> flt_src = pthr->src;
>> bnd_dst = peer->outbuf + peer->outbuf_size;
>> flt_dst = peer->outbuf;
>> + bnd_phys_addr = peer->out_phys_addr + peer->outbuf_size;
>> + flt_phys_addr = peer->out_phys_addr;
>>
>> pthr->duration = ktime_get();
>>
>> /* Copied field is cleared on test launch stage */
>> while (pthr->copied < total_size) {
>> - ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size);
>> + ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size,
>> + flt_phys_addr);
>> if (ret) {
>> dev_err(&perf->ntb->dev, "%d: Got error %d on test\n",
>> pthr->tidx, ret);
>> @@ -925,8 +931,15 @@ static int perf_run_test(struct perf_thread *pthr)
>>
>> flt_dst += chunk_size;
>> flt_src += chunk_size;
>> - if (flt_dst >= bnd_dst || flt_dst < peer->outbuf) {
>> + flt_phys_addr += chunk_size;
>> +
>> + if (flt_dst >= bnd_dst ||
>> + flt_dst < peer->outbuf ||
>> + flt_phys_addr >= bnd_phys_addr ||
>> + flt_phys_addr < peer->out_phys_addr) {
>> +
>> flt_dst = peer->outbuf;
>> + flt_phys_addr = peer->out_phys_addr;
>> flt_src = pthr->src;
>> }
>>
>> @@ -1195,6 +1208,9 @@ static ssize_t perf_dbgfs_read_info(struct file *filep, char __user *ubuf,
>> "\tOut buffer addr 0x%pK\n", peer->outbuf);
>>
>> pos += scnprintf(buf + pos, buf_size - pos,
>> + "\tOut buff phys addr %pa[p]\n", &peer->out_phys_addr);
>> +
>> + pos += scnprintf(buf + pos, buf_size - pos,
>> "\tOut buffer size %pa\n", &peer->outbuf_size);
>>
>> pos += scnprintf(buf + pos, buf_size - pos,
>> @@ -1388,6 +1404,8 @@ static int perf_setup_peer_mw(struct perf_peer *peer)
>> if (!peer->outbuf)
>> return -ENOMEM;
>>
>> + peer->out_phys_addr = phys_addr;
>> +
>> if (max_mw_size && peer->outbuf_size > max_mw_size) {
>> peer->outbuf_size = max_mw_size;
>> dev_warn(&peer->perf->ntb->dev,
>>
On 2019-11-11 12:51 a.m., Jiasen Lin wrote:
>
>
> On 2019/11/9 1:04, Logan Gunthorpe wrote:
>>
>>
>> On 2019-11-06 8:38 p.m., Jiasen Lin wrote:
>>> peer->outbuf is a virtual address which is get by ioremap, it can not
>>> be converted to a physical address by virt_to_page and page_to_phys.
>>> This conversion will result in DMA error, because the destination address
>>> which is converted by page_to_phys is invalid.
>>
>> Hmm, yes, ntb_perf is obviously wrong. I never noticed this, how did
>> this ever work?
>>
>
> The default value of use_dma which is used to enable DMA engine to
> measure NTB performance is zero, in other words, DMA engine is not used
> by default. Thus, olny memcpy_toio is called in perf_copy_chunk and not
> trigger this bug.
>
> If we install driver with specified dma-enabled flag like this:
> insmod ntb_perf.ko use_dma=1, this issue will be triggered.
I've definitely tested with use_dma=1 in the past. But it looks like it
was broken by this problematic commit and I must have never personally
run the DMA tests after it:
5648e56d03fa ("NTB: ntb_perf: Add full multi-port NTB API support")
So it's probably worth adding a fixes tag to your patch.
>>> We Save the physical address in perf_setup_peer_mw, it is MMIO address
>>> of NTB BARx. Then fill the destination address of DMA descriptor with
>>> this physical address to guarantee that the address of memory write
>>> requests fall in memory window of NBT BARx.
>>
>> Using the physical address directly is also wrong and will not work in
>> the presence of an IOMMU. You should use dma_map_resource() for this.
>> See what was done in ntb_transport[1].
>>
>
> Yes, my mistake. I will modify the code according to your suggestion and
> test it on AMD and HYGON platforms with the IOMMU enabled. Maybe the
> following patches are relied on, when IOMMU is enabled on AMD and HYGON
> plartforms.
>
> https://lore.kernel.org/patchwork/cover/1143155/
> https://lore.kernel.org/patchwork/patch/1143156/
> https://lore.kernel.org/patchwork/patch/1143157/
Thanks!
Logan