Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com --- net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv; + size_t size;
if (!binding) return -EINVAL; @@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
+ /* Validate that the underlying dmabuf has enough memory to satisfy + * requested pool size. + */ + size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT; + if (size < pool->p.pool_size) { + pr_warn("%s: Insufficient dmabuf memory (%zu pages) to satisfy pool_size (%u pages)\n", + __func__, size, pool->p.pool_size); + return -ENOMEM; + } + net_devmem_dmabuf_binding_get(binding); return 0; }
netdev_bind_rx takes ownership of the queue array passed as parameter and frees it, so a queue array buffer cannot be reused across multiple netdev_bind_rx calls.
This commit fixes that by always passing in a newly created queue array to all netdev_bind_rx calls in ncdevmem.
Fixes: 85585b4bc8d8 ("selftests: add ncdevmem, netcat for devmem TCP") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com --- .../selftests/drivers/net/hw/ncdevmem.c | 55 ++++++++----------- 1 file changed, 22 insertions(+), 33 deletions(-)
diff --git a/tools/testing/selftests/drivers/net/hw/ncdevmem.c b/tools/testing/selftests/drivers/net/hw/ncdevmem.c index 2bf14ac2b8c6..9d48004ff1a1 100644 --- a/tools/testing/selftests/drivers/net/hw/ncdevmem.c +++ b/tools/testing/selftests/drivers/net/hw/ncdevmem.c @@ -431,6 +431,22 @@ static int parse_address(const char *str, int port, struct sockaddr_in6 *sin6) return 0; }
+static struct netdev_queue_id *create_queues(void) +{ + struct netdev_queue_id *queues; + size_t i = 0; + + queues = calloc(num_queues, sizeof(*queues)); + for (i = 0; i < num_queues; i++) { + queues[i]._present.type = 1; + queues[i]._present.id = 1; + queues[i].type = NETDEV_QUEUE_TYPE_RX; + queues[i].id = start_queue + i; + } + + return queues; +} + int do_server(struct memory_buffer *mem) { char ctrl_data[sizeof(int) * 20000]; @@ -448,7 +464,6 @@ int do_server(struct memory_buffer *mem) char buffer[256]; int socket_fd; int client_fd; - size_t i = 0; int ret;
ret = parse_address(server_ip, atoi(port), &server_sin); @@ -471,16 +486,7 @@ int do_server(struct memory_buffer *mem)
sleep(1);
- queues = malloc(sizeof(*queues) * num_queues); - - for (i = 0; i < num_queues; i++) { - queues[i]._present.type = 1; - queues[i]._present.id = 1; - queues[i].type = NETDEV_QUEUE_TYPE_RX; - queues[i].id = start_queue + i; - } - - if (bind_rx_queue(ifindex, mem->fd, queues, num_queues, &ys)) + if (bind_rx_queue(ifindex, mem->fd, create_queues(), num_queues, &ys)) error(1, 0, "Failed to bind\n");
tmp_mem = malloc(mem->size); @@ -545,7 +551,6 @@ int do_server(struct memory_buffer *mem) goto cleanup; }
- i++; for (cm = CMSG_FIRSTHDR(&msg); cm; cm = CMSG_NXTHDR(&msg, cm)) { if (cm->cmsg_level != SOL_SOCKET || (cm->cmsg_type != SCM_DEVMEM_DMABUF && @@ -630,10 +635,8 @@ int do_server(struct memory_buffer *mem)
void run_devmem_tests(void) { - struct netdev_queue_id *queues; struct memory_buffer *mem; struct ynl_sock *ys; - size_t i = 0;
mem = provider->alloc(getpagesize() * NUM_PAGES);
@@ -641,38 +644,24 @@ void run_devmem_tests(void) if (configure_rss()) error(1, 0, "rss error\n");
- queues = calloc(num_queues, sizeof(*queues)); - if (configure_headersplit(1)) error(1, 0, "Failed to configure header split\n");
- if (!bind_rx_queue(ifindex, mem->fd, queues, num_queues, &ys)) + if (!bind_rx_queue(ifindex, mem->fd, + calloc(num_queues, sizeof(struct netdev_queue_id)), + num_queues, &ys)) error(1, 0, "Binding empty queues array should have failed\n");
- for (i = 0; i < num_queues; i++) { - queues[i]._present.type = 1; - queues[i]._present.id = 1; - queues[i].type = NETDEV_QUEUE_TYPE_RX; - queues[i].id = start_queue + i; - } - if (configure_headersplit(0)) error(1, 0, "Failed to configure header split\n");
- if (!bind_rx_queue(ifindex, mem->fd, queues, num_queues, &ys)) + if (!bind_rx_queue(ifindex, mem->fd, create_queues(), num_queues, &ys)) error(1, 0, "Configure dmabuf with header split off should have failed\n");
if (configure_headersplit(1)) error(1, 0, "Failed to configure header split\n");
- for (i = 0; i < num_queues; i++) { - queues[i]._present.type = 1; - queues[i]._present.id = 1; - queues[i].type = NETDEV_QUEUE_TYPE_RX; - queues[i].id = start_queue + i; - } - - if (bind_rx_queue(ifindex, mem->fd, queues, num_queues, &ys)) + if (bind_rx_queue(ifindex, mem->fd, create_queues(), num_queues, &ys)) error(1, 0, "Failed to bind\n");
/* Deactivating a bound queue should not be legal */
On 04/23, Cosmin Ratiu wrote:
netdev_bind_rx takes ownership of the queue array passed as parameter and frees it, so a queue array buffer cannot be reused across multiple netdev_bind_rx calls.
This commit fixes that by always passing in a newly created queue array to all netdev_bind_rx calls in ncdevmem.
Fixes: 85585b4bc8d8 ("selftests: add ncdevmem, netcat for devmem TCP") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
Acked-by: Stanislav Fomichev sdf@fomichev.me
On Wed, Apr 23, 2025 at 9:00 AM Cosmin Ratiu cratiu@nvidia.com wrote:
netdev_bind_rx takes ownership of the queue array passed as parameter and frees it, so a queue array buffer cannot be reused across multiple netdev_bind_rx calls.
This commit fixes that by always passing in a newly created queue array to all netdev_bind_rx calls in ncdevmem.
Fixes: 85585b4bc8d8 ("selftests: add ncdevmem, netcat for devmem TCP") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
Thank you!
Reviewed-by: Mina Almasry almasrymina@google.com
On 04/23, Cosmin Ratiu wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
- size_t size;
if (!binding) return -EINVAL; @@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
- /* Validate that the underlying dmabuf has enough memory to satisfy
* requested pool size.*/
I think it's useful to have a check, but note that this check is in no way a guarantee that the genpool has enough capacity. We can use the same binding on multiple queues... Can you expand the comment a bit to explain that it's more of a sanity check than a guarantee?
- size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT;
- if (size < pool->p.pool_size) {
pr_warn("%s: Insufficient dmabuf memory (%zu pages) to satisfy pool_size (%u pages)\n",
Let's print the sizes in bytes? We might have order>0 pages soon in the pp: https://lore.kernel.org/netdev/20250421222827.283737-1-kuba@kernel.org/T/#t
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
size_t size; if (!binding) return -EINVAL;@@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
/* Validate that the underlying dmabuf has enough memory to satisfy* requested pool size.*/size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT;if (size < pool->p.pool_size) {
pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
pr_warn("%s: Insufficient dmabuf memory (%zu pages) to satisfy pool_size (%u pages)\n",__func__, size, pool->p.pool_size);return -ENOMEM;
In general I think maybe printing an error/warn to dmesg to warn the user that this is a weird configuration may be fine, but return -ENOMEM? I don't think this should be an unsupported configuration and I don't think checking against p.pool_size guarantees anything.
}net_devmem_dmabuf_binding_get(binding); return 0;}
2.45.0
On 04/23, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
size_t size; if (!binding) return -EINVAL;@@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
/* Validate that the underlying dmabuf has enough memory to satisfy* requested pool size.*/size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT;if (size < pool->p.pool_size) {pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I read this check more as "is there enough chunks in the binding to fully fill in the page pool". User controls the size of rx ring which controls the size of the page pool which somewhat dictates the minimal size of the binding (maybe). So it's more of a sanity check.
Maybe having better defaults in ncdevmem would've been a better option? It allocates (16000*4096) bytes (slightly less than 64MB, why? to fit into default /sys/module/udmabuf/parameters/size_limit_mb?) and on my setup PP wants to get 64MB at least..
On Wed, Apr 23, 2025 at 1:15 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/23, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
size_t size; if (!binding) return -EINVAL;@@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
/* Validate that the underlying dmabuf has enough memory to satisfy* requested pool size.*/size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT;if (size < pool->p.pool_size) {pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I read this check more as "is there enough chunks in the binding to fully fill in the page pool". User controls the size of rx ring
Only on drivers that support ethtool -G, and where it will let you configure -G to what you want.
which controls the size of the page pool which somewhat dictates the minimal size of the binding (maybe).
See the test I ran in the other thread. Seems at least GVE is fine with dmabuf size < ring size. I don't know what other drivers do, but generally speaking I think specific driver limitations should not limit what others can do with their drivers. Sure for the GPU mem applications you're probably looking at the dmabufs are huge and supporting small dmabufs is not a concern, but someone somewhere may want to run with 1 MB dmabuf for some use case and if their driver is fine with it, core should not prevent it, I think.
So it's more of a sanity check.
Maybe having better defaults in ncdevmem would've been a better option? It allocates (16000*4096) bytes (slightly less than 64MB, why? to fit into default /sys/module/udmabuf/parameters/size_limit_mb?) and on my setup PP wants to get 64MB at least..
Yeah, udmabuf has a limitation that it only supports 64MB max size last I looked.
I added devmem TCP support with udmabuf selftests to demonstrate that the feature is non-proprietary, not to advertise that devmem tcp + udmabuf is a great combination. udmabuf is actually terrible for devmem TCP. The 64MB limit is way too small for anyone to do anything performant on it and by dmaing into host memory you lose many of the benefits of devmem TCP (lower mem bw + pcie bw utilization).
If you're running real experiments with devmem TCP I suggest moving to real dmabufs as soon as possible, or at least hack udmabuf to give you large sizes. We've open sourced our production devmem TCP userspace:
https://github.com/google/tcpgpudmarxd https://github.com/google/nccl-plugin-gpudirecttcpx
Porting it to upstream APIs + your dmabuf provider will have you run much more interesting tests than anything you do with udmabuf I think, unless you hack the udmabuf size.
On 04/24, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 1:15 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/23, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
size_t size; if (!binding) return -EINVAL;@@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
/* Validate that the underlying dmabuf has enough memory to satisfy* requested pool size.*/size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT;if (size < pool->p.pool_size) {pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I read this check more as "is there enough chunks in the binding to fully fill in the page pool". User controls the size of rx ring
Only on drivers that support ethtool -G, and where it will let you configure -G to what you want.
gve is the minority here, any major nic (brcm/mlx/intel) supports resizing the rings.
which controls the size of the page pool which somewhat dictates the minimal size of the binding (maybe).
See the test I ran in the other thread. Seems at least GVE is fine with dmabuf size < ring size. I don't know what other drivers do, but generally speaking I think specific driver limitations should not limit what others can do with their drivers. Sure for the GPU mem applications you're probably looking at the dmabufs are huge and supporting small dmabufs is not a concern, but someone somewhere may want to run with 1 MB dmabuf for some use case and if their driver is fine with it, core should not prevent it, I think.
So it's more of a sanity check.
Maybe having better defaults in ncdevmem would've been a better option? It allocates (16000*4096) bytes (slightly less than 64MB, why? to fit into default /sys/module/udmabuf/parameters/size_limit_mb?) and on my setup PP wants to get 64MB at least..
Yeah, udmabuf has a limitation that it only supports 64MB max size last I looked.
We can use /sys/module/udmabuf/parameters/size_limit_mb to allocate more than 64MB, ncdevmem can change it. Or warn the user similar to what kperf does: https://github.com/facebookexperimental/kperf/blob/main/devmem.c#L308
So either having a kernel warn or tuning 63MB up to something sensible (1G?) should prevent people from going through the pain..
I added devmem TCP support with udmabuf selftests to demonstrate that the feature is non-proprietary, not to advertise that devmem tcp + udmabuf is a great combination. udmabuf is actually terrible for devmem TCP. The 64MB limit is way too small for anyone to do anything performant on it and by dmaing into host memory you lose many of the benefits of devmem TCP (lower mem bw + pcie bw utilization).
It would still be nice to have a udmabuf as a properly supported option. This can drive the UAPI performance conversions: for example, comparing existing tcp rx zerocopy vs MSG_SOCK_DEVMEM.. So let's not completely dismiss it. We've played internally with doing 2MB udmabuf huge-pages, might post it at some point..
If you're running real experiments with devmem TCP I suggest moving to real dmabufs as soon as possible, or at least hack udmabuf to give you large sizes. We've open sourced our production devmem TCP userspace:
https://github.com/google/tcpgpudmarxd https://github.com/google/nccl-plugin-gpudirecttcpx
Porting it to upstream APIs + your dmabuf provider will have you run much more interesting tests than anything you do with udmabuf I think, unless you hack the udmabuf size.
I found these a bit too late, so I reimplemented the plugin over upstream APIs :-[ Plus, you yourself have acked [0], guess why I sent this patch :-D Once the tx part is accepted, we'll upstream kperf cuda support as well..
0: https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?...
On Thu, Apr 24, 2025 at 3:10 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/24, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 1:15 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/23, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
size_t size; if (!binding) return -EINVAL;@@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
/* Validate that the underlying dmabuf has enough memory to satisfy* requested pool size.*/size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT;if (size < pool->p.pool_size) {pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I read this check more as "is there enough chunks in the binding to fully fill in the page pool". User controls the size of rx ring
Only on drivers that support ethtool -G, and where it will let you configure -G to what you want.
gve is the minority here, any major nic (brcm/mlx/intel) supports resizing the rings.
GVE supports resizing rings. Other drivers may not. Even on drivers that support resizing rings. Some users may have a use case for a dmabuf smaller than the minimum ring size their driver accepts.
which controls the size of the page pool which somewhat dictates the minimal size of the binding (maybe).
See the test I ran in the other thread. Seems at least GVE is fine with dmabuf size < ring size. I don't know what other drivers do, but generally speaking I think specific driver limitations should not limit what others can do with their drivers. Sure for the GPU mem applications you're probably looking at the dmabufs are huge and supporting small dmabufs is not a concern, but someone somewhere may want to run with 1 MB dmabuf for some use case and if their driver is fine with it, core should not prevent it, I think.
So it's more of a sanity check.
Maybe having better defaults in ncdevmem would've been a better option? It allocates (16000*4096) bytes (slightly less than 64MB, why? to fit into default /sys/module/udmabuf/parameters/size_limit_mb?) and on my setup PP wants to get 64MB at least..
Yeah, udmabuf has a limitation that it only supports 64MB max size last I looked.
We can use /sys/module/udmabuf/parameters/size_limit_mb to allocate more than 64MB, ncdevmem can change it.
The udmabuf limit is hardcoded, in udmabuf.c or configured on module load, and ncdevmem doesn't load udmabuf. I guess it could be changed to that, but currently ncdevmem works with CONFIG_UDMABUF=y
Or warn the user similar to what kperf does: https://github.com/facebookexperimental/kperf/blob/main/devmem.c#L308
So either having a kernel warn or tuning 63MB up to something sensible (1G?) should prevent people from going through the pain..
Agreed with both. Another option is updating the devmem.rst docs:
"Some drivers may struggle to run devmem TCP when the dmabuf size is too small, especially when it's smaller than the number of rx ring slots. Look for this warning in dmesg." etc.
But I don't see the need to outright disable this "dmabuf size < ring size" use case for everyone to solve this.
I added devmem TCP support with udmabuf selftests to demonstrate that the feature is non-proprietary, not to advertise that devmem tcp + udmabuf is a great combination. udmabuf is actually terrible for devmem TCP. The 64MB limit is way too small for anyone to do anything performant on it and by dmaing into host memory you lose many of the benefits of devmem TCP (lower mem bw + pcie bw utilization).
It would still be nice to have a udmabuf as a properly supported option. This can drive the UAPI performance conversions: for example, comparing existing tcp rx zerocopy vs MSG_SOCK_DEVMEM.. So let's not completely dismiss it. We've played internally with doing 2MB udmabuf huge-pages, might post it at some point..
If you're running real experiments with devmem TCP I suggest moving to real dmabufs as soon as possible, or at least hack udmabuf to give you large sizes. We've open sourced our production devmem TCP userspace:
https://github.com/google/tcpgpudmarxd https://github.com/google/nccl-plugin-gpudirecttcpx
Porting it to upstream APIs + your dmabuf provider will have you run much more interesting tests than anything you do with udmabuf I think, unless you hack the udmabuf size.
I found these a bit too late, so I reimplemented the plugin over upstream APIs :-[
Oh, where? Is it open source?
Plus, you yourself have acked [0], guess why I sent this patch :-D Once the tx part is accepted, we'll upstream kperf cuda support as well..
Cool!
0: https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?...
On 04/24, Mina Almasry wrote:
On Thu, Apr 24, 2025 at 3:10 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/24, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 1:15 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/23, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
This commit adds a check at dmabuf pool init time that compares the amount of memory in the underlying chunk pool (configured by the user space application providing dmabuf memory) with the desired pool size (previously set by the driver) and fails with an error message if chunk memory isn't enough.
Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") Signed-off-by: Cosmin Ratiu cratiu@nvidia.com
net/core/devmem.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/net/core/devmem.c b/net/core/devmem.c index 6e27a47d0493..651cd55ebb28 100644 --- a/net/core/devmem.c +++ b/net/core/devmem.c @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, int mp_dmabuf_devmem_init(struct page_pool *pool) { struct net_devmem_dmabuf_binding *binding = pool->mp_priv;
size_t size; if (!binding) return -EINVAL;@@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) if (pool->p.order != 0) return -E2BIG;
/* Validate that the underlying dmabuf has enough memory to satisfy* requested pool size.*/size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT;if (size < pool->p.pool_size) {pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I read this check more as "is there enough chunks in the binding to fully fill in the page pool". User controls the size of rx ring
Only on drivers that support ethtool -G, and where it will let you configure -G to what you want.
gve is the minority here, any major nic (brcm/mlx/intel) supports resizing the rings.
GVE supports resizing rings. Other drivers may not. Even on drivers that support resizing rings. Some users may have a use case for a dmabuf smaller than the minimum ring size their driver accepts.
which controls the size of the page pool which somewhat dictates the minimal size of the binding (maybe).
See the test I ran in the other thread. Seems at least GVE is fine with dmabuf size < ring size. I don't know what other drivers do, but generally speaking I think specific driver limitations should not limit what others can do with their drivers. Sure for the GPU mem applications you're probably looking at the dmabufs are huge and supporting small dmabufs is not a concern, but someone somewhere may want to run with 1 MB dmabuf for some use case and if their driver is fine with it, core should not prevent it, I think.
So it's more of a sanity check.
Maybe having better defaults in ncdevmem would've been a better option? It allocates (16000*4096) bytes (slightly less than 64MB, why? to fit into default /sys/module/udmabuf/parameters/size_limit_mb?) and on my setup PP wants to get 64MB at least..
Yeah, udmabuf has a limitation that it only supports 64MB max size last I looked.
We can use /sys/module/udmabuf/parameters/size_limit_mb to allocate more than 64MB, ncdevmem can change it.
The udmabuf limit is hardcoded, in udmabuf.c or configured on module load, and ncdevmem doesn't load udmabuf. I guess it could be changed to that, but currently ncdevmem works with CONFIG_UDMABUF=y
You don't need to load/reload the module to change module params:
# id uid=0(root) gid=0(root) groups=0(root),1(bin),2(daemon),3(sys) # cat /sys/module/udmabuf/parameters/size_limit_mb 64 # echo 128 > /sys/module/udmabuf/parameters/size_limit_mb # cat /sys/module/udmabuf/parameters/size_limit_mb 128
Or warn the user similar to what kperf does: https://github.com/facebookexperimental/kperf/blob/main/devmem.c#L308
So either having a kernel warn or tuning 63MB up to something sensible (1G?) should prevent people from going through the pain..
Agreed with both. Another option is updating the devmem.rst docs:
"Some drivers may struggle to run devmem TCP when the dmabuf size is too small, especially when it's smaller than the number of rx ring slots. Look for this warning in dmesg." etc.
But I don't see the need to outright disable this "dmabuf size < ring size" use case for everyone to solve this.
Agreed. The fact that mlx5 has issues with small pp should be fixed, I'm not arguing with that. I'm trying to understand whether giving a hint to the user about dmabuf < pp size is helpful or not (because it will most likely will lead to poor perf, which is the main point of devmem).
I added devmem TCP support with udmabuf selftests to demonstrate that the feature is non-proprietary, not to advertise that devmem tcp + udmabuf is a great combination. udmabuf is actually terrible for devmem TCP. The 64MB limit is way too small for anyone to do anything performant on it and by dmaing into host memory you lose many of the benefits of devmem TCP (lower mem bw + pcie bw utilization).
It would still be nice to have a udmabuf as a properly supported option. This can drive the UAPI performance conversions: for example, comparing existing tcp rx zerocopy vs MSG_SOCK_DEVMEM.. So let's not completely dismiss it. We've played internally with doing 2MB udmabuf huge-pages, might post it at some point..
If you're running real experiments with devmem TCP I suggest moving to real dmabufs as soon as possible, or at least hack udmabuf to give you large sizes. We've open sourced our production devmem TCP userspace:
https://github.com/google/tcpgpudmarxd https://github.com/google/nccl-plugin-gpudirecttcpx
Porting it to upstream APIs + your dmabuf provider will have you run much more interesting tests than anything you do with udmabuf I think, unless you hack the udmabuf size.
I found these a bit too late, so I reimplemented the plugin over upstream APIs :-[
Oh, where? Is it open source?
No, but depending on where we end up I think it should be possible to make it open source. The is no secret sauce in there..
On Thu, Apr 24, 2025 at 3:40 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/24, Mina Almasry wrote:
On Thu, Apr 24, 2025 at 3:10 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/24, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 1:15 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/23, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote: > > Drivers that are told to allocate RX buffers from pools of DMA memory > should have enough memory in the pool to satisfy projected allocation > requests (a function of ring size, MTU & other parameters). If there's > not enough memory, RX ring refill might fail later at inconvenient times > (e.g. during NAPI poll). >
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
> This commit adds a check at dmabuf pool init time that compares the > amount of memory in the underlying chunk pool (configured by the user > space application providing dmabuf memory) with the desired pool size > (previously set by the driver) and fails with an error message if chunk > memory isn't enough. > > Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") > Signed-off-by: Cosmin Ratiu cratiu@nvidia.com > --- > net/core/devmem.c | 11 +++++++++++ > 1 file changed, 11 insertions(+) > > diff --git a/net/core/devmem.c b/net/core/devmem.c > index 6e27a47d0493..651cd55ebb28 100644 > --- a/net/core/devmem.c > +++ b/net/core/devmem.c > @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, > int mp_dmabuf_devmem_init(struct page_pool *pool) > { > struct net_devmem_dmabuf_binding *binding = pool->mp_priv; > + size_t size; > > if (!binding) > return -EINVAL; > @@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) > if (pool->p.order != 0) > return -E2BIG; > > + /* Validate that the underlying dmabuf has enough memory to satisfy > + * requested pool size. > + */ > + size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT; > + if (size < pool->p.pool_size) {
pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I read this check more as "is there enough chunks in the binding to fully fill in the page pool". User controls the size of rx ring
Only on drivers that support ethtool -G, and where it will let you configure -G to what you want.
gve is the minority here, any major nic (brcm/mlx/intel) supports resizing the rings.
GVE supports resizing rings. Other drivers may not. Even on drivers that support resizing rings. Some users may have a use case for a dmabuf smaller than the minimum ring size their driver accepts.
which controls the size of the page pool which somewhat dictates the minimal size of the binding (maybe).
See the test I ran in the other thread. Seems at least GVE is fine with dmabuf size < ring size. I don't know what other drivers do, but generally speaking I think specific driver limitations should not limit what others can do with their drivers. Sure for the GPU mem applications you're probably looking at the dmabufs are huge and supporting small dmabufs is not a concern, but someone somewhere may want to run with 1 MB dmabuf for some use case and if their driver is fine with it, core should not prevent it, I think.
So it's more of a sanity check.
Maybe having better defaults in ncdevmem would've been a better option? It allocates (16000*4096) bytes (slightly less than 64MB, why? to fit into default /sys/module/udmabuf/parameters/size_limit_mb?) and on my setup PP wants to get 64MB at least..
Yeah, udmabuf has a limitation that it only supports 64MB max size last I looked.
We can use /sys/module/udmabuf/parameters/size_limit_mb to allocate more than 64MB, ncdevmem can change it.
The udmabuf limit is hardcoded, in udmabuf.c or configured on module load, and ncdevmem doesn't load udmabuf. I guess it could be changed to that, but currently ncdevmem works with CONFIG_UDMABUF=y
You don't need to load/reload the module to change module params:
# id uid=0(root) gid=0(root) groups=0(root),1(bin),2(daemon),3(sys) # cat /sys/module/udmabuf/parameters/size_limit_mb 64 # echo 128 > /sys/module/udmabuf/parameters/size_limit_mb # cat /sys/module/udmabuf/parameters/size_limit_mb 128
Today I learned! Thanks!
I will put it on my todo list to make ncdevmem force a larger limit to udmabuf and allocate a larger dmabuf by default. Technically any dmabuf should be supported, but by default I think probably ncdevmem should use a .5 GB -> 1GB dmabuf that is more common for these GPU applications or something. There could be an option as well for folks to test their driver against smaller dmabufs.
On 04/24, Mina Almasry wrote:
On Thu, Apr 24, 2025 at 3:40 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/24, Mina Almasry wrote:
On Thu, Apr 24, 2025 at 3:10 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/24, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 1:15 PM Stanislav Fomichev stfomichev@gmail.com wrote:
On 04/23, Mina Almasry wrote: > On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote: > > > > Drivers that are told to allocate RX buffers from pools of DMA memory > > should have enough memory in the pool to satisfy projected allocation > > requests (a function of ring size, MTU & other parameters). If there's > > not enough memory, RX ring refill might fail later at inconvenient times > > (e.g. during NAPI poll). > > > > My understanding is that if the RX ring refill fails, the driver will > post the buffers it was able to allocate data for, and will not post > other buffers. So it will run with a degraded performance but nothing > overly bad should happen. This should be the same behavior if the > machine is under memory pressure. > > In general I don't know about this change. If the user wants to use > very small dmabufs, they should be able to, without going through > hoops reducing the number of rx ring slots the driver has (if it > supports configuring that). > > I think maybe printing an error or warning that the dmabuf is too > small for the pool_size may be fine. But outright failing this > configuration? I don't think so. > > > This commit adds a check at dmabuf pool init time that compares the > > amount of memory in the underlying chunk pool (configured by the user > > space application providing dmabuf memory) with the desired pool size > > (previously set by the driver) and fails with an error message if chunk > > memory isn't enough. > > > > Fixes: 0f9214046893 ("memory-provider: dmabuf devmem memory provider") > > Signed-off-by: Cosmin Ratiu cratiu@nvidia.com > > --- > > net/core/devmem.c | 11 +++++++++++ > > 1 file changed, 11 insertions(+) > > > > diff --git a/net/core/devmem.c b/net/core/devmem.c > > index 6e27a47d0493..651cd55ebb28 100644 > > --- a/net/core/devmem.c > > +++ b/net/core/devmem.c > > @@ -299,6 +299,7 @@ net_devmem_bind_dmabuf(struct net_device *dev, unsigned int dmabuf_fd, > > int mp_dmabuf_devmem_init(struct page_pool *pool) > > { > > struct net_devmem_dmabuf_binding *binding = pool->mp_priv; > > + size_t size; > > > > if (!binding) > > return -EINVAL; > > @@ -312,6 +313,16 @@ int mp_dmabuf_devmem_init(struct page_pool *pool) > > if (pool->p.order != 0) > > return -E2BIG; > > > > + /* Validate that the underlying dmabuf has enough memory to satisfy > > + * requested pool size. > > + */ > > + size = gen_pool_size(binding->chunk_pool) >> PAGE_SHIFT; > > + if (size < pool->p.pool_size) { > > pool_size seems to be the number of ptr_ring slots in the page_pool, > not some upper or lower bound on the amount of memory the page_pool > can provide. So this check seems useless? The page_pool can still not > provide this amount of memory with dmabuf (if the netmems aren't being > recycled fast enough) or with normal memory (under memory pressure).
I read this check more as "is there enough chunks in the binding to fully fill in the page pool". User controls the size of rx ring
Only on drivers that support ethtool -G, and where it will let you configure -G to what you want.
gve is the minority here, any major nic (brcm/mlx/intel) supports resizing the rings.
GVE supports resizing rings. Other drivers may not. Even on drivers that support resizing rings. Some users may have a use case for a dmabuf smaller than the minimum ring size their driver accepts.
which controls the size of the page pool which somewhat dictates the minimal size of the binding (maybe).
See the test I ran in the other thread. Seems at least GVE is fine with dmabuf size < ring size. I don't know what other drivers do, but generally speaking I think specific driver limitations should not limit what others can do with their drivers. Sure for the GPU mem applications you're probably looking at the dmabufs are huge and supporting small dmabufs is not a concern, but someone somewhere may want to run with 1 MB dmabuf for some use case and if their driver is fine with it, core should not prevent it, I think.
So it's more of a sanity check.
Maybe having better defaults in ncdevmem would've been a better option? It allocates (16000*4096) bytes (slightly less than 64MB, why? to fit into default /sys/module/udmabuf/parameters/size_limit_mb?) and on my setup PP wants to get 64MB at least..
Yeah, udmabuf has a limitation that it only supports 64MB max size last I looked.
We can use /sys/module/udmabuf/parameters/size_limit_mb to allocate more than 64MB, ncdevmem can change it.
The udmabuf limit is hardcoded, in udmabuf.c or configured on module load, and ncdevmem doesn't load udmabuf. I guess it could be changed to that, but currently ncdevmem works with CONFIG_UDMABUF=y
You don't need to load/reload the module to change module params:
# id uid=0(root) gid=0(root) groups=0(root),1(bin),2(daemon),3(sys) # cat /sys/module/udmabuf/parameters/size_limit_mb 64 # echo 128 > /sys/module/udmabuf/parameters/size_limit_mb # cat /sys/module/udmabuf/parameters/size_limit_mb 128
Today I learned! Thanks!
I will put it on my todo list to make ncdevmem force a larger limit to
Or we can ask Cosmin to send something out? Since he's already looking into the buffer sizes..
On Wed, 2025-04-23 at 10:30 -0700, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
What motivated this change was a failure in how mlx5 refills rings today. For efficiency, ring refill triggered by NAPI polling only releases old buffers just before allocating new ones so effectively has a built-in assumption that the ring can be filled. Commit 4c2a13236807 ("net/mlx5e: RX, Defer page release in striding rq for better recycling") is an interesting read here.
For more details, see the do{ }while loop in mlx5e_post_rx_mpwqes, where mlx5e_free_rx_mpwqe then mlx5e_alloc_rx_mpwqe are called to free the old buffer and reallocate a new one at the same position. This has excellent cache-locality and the pages returned to the pool will be reused by the new descriptor.
The bug in mlx5 is that with a large MTU & ring size, the ring cannot be fully populated with rx descriptors because the pool doesn't have enough memory, but there's no memory released back to the pool for new ones. Eventually, rx descriptors are exhausted and traffic stops.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
For regular memory-backed page pools, there's no hard limit of how big they can become (except available physical memory), so this issue was not seen before.
I didn't look at other drivers yet, but is it expected that drivers operate with incompletely filled rings? I assumed that since the user configured a specified ring size, it expects drivers to be able to use that size and not silently operate in degraded mode, with a smaller ring size.
If you think drivers should work in degraded mode, we can look at improving the ring population code to work better in this scenario.
pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I think pool_size is usually set by drivers based on their params, and it's the max size of pool->ring. The opportunistic check I added compares this demand with the supply (available chunk memory) and fails this config based on the assumption that there should be enough memory in the pool to satisfy driver needs.
Please let me know your thoughts and how to proceed.
Cosmin.
On Thu, Apr 24, 2025 at 1:47 AM Cosmin Ratiu cratiu@nvidia.com wrote:
On Wed, 2025-04-23 at 10:30 -0700, Mina Almasry wrote:
On Wed, Apr 23, 2025 at 9:03 AM Cosmin Ratiu cratiu@nvidia.com wrote:
Drivers that are told to allocate RX buffers from pools of DMA memory should have enough memory in the pool to satisfy projected allocation requests (a function of ring size, MTU & other parameters). If there's not enough memory, RX ring refill might fail later at inconvenient times (e.g. during NAPI poll).
My understanding is that if the RX ring refill fails, the driver will post the buffers it was able to allocate data for, and will not post other buffers. So it will run with a degraded performance but nothing overly bad should happen. This should be the same behavior if the machine is under memory pressure.
What motivated this change was a failure in how mlx5 refills rings today. For efficiency, ring refill triggered by NAPI polling only releases old buffers just before allocating new ones so effectively has a built-in assumption that the ring can be filled. Commit 4c2a13236807 ("net/mlx5e: RX, Defer page release in striding rq for better recycling") is an interesting read here.
For more details, see the do{ }while loop in mlx5e_post_rx_mpwqes, where mlx5e_free_rx_mpwqe then mlx5e_alloc_rx_mpwqe are called to free the old buffer and reallocate a new one at the same position. This has excellent cache-locality and the pages returned to the pool will be reused by the new descriptor.
Thanks for the detailed explanation. These seem like a clever optimization.
The bug in mlx5 is that with a large MTU & ring size, the ring cannot be fully populated with rx descriptors because the pool doesn't have enough memory, but there's no memory released back to the pool for new ones. Eventually, rx descriptors are exhausted and traffic stops.
In general I don't know about this change. If the user wants to use very small dmabufs, they should be able to, without going through hoops reducing the number of rx ring slots the driver has (if it supports configuring that).
I think maybe printing an error or warning that the dmabuf is too small for the pool_size may be fine. But outright failing this configuration? I don't think so.
For regular memory-backed page pools, there's no hard limit of how big they can become (except available physical memory), so this issue was not seen before.
I didn't look at other drivers yet, but is it expected that drivers operate with incompletely filled rings? I assumed that since the user configured a specified ring size, it expects drivers to be able to use that size and not silently operate in degraded mode, with a smaller ring size.
If you think drivers should work in degraded mode, we can look at improving the ring population code to work better in this scenario.
You're probably a bigger expert to me on what drivers should do in general, but yes, this seems like an mlx5 limitation, not a general limitation to all drivers. GVE for example, I think, has a host of optimization for memory pressure scenarios that makes it resilient to this. I ran this test:
mina-3 /home/almasrymina_google_com # ethtool -g eth0 Ring parameters for eth0: ... Current hardware settings: ... RX: 1024
Then hacked ncdevmem to only provide a 64 page udmabuf:
diff --git a/tools/testing/selftests/drivers/net/hw/ncdevmem.c b/tools/testing/selftests/drivers/net/hw/ncdevmem.c index 1f9fb0b1cb811..6de64f7680241 100644 --- a/tools/testing/selftests/drivers/net/hw/ncdevmem.c +++ b/tools/testing/selftests/drivers/net/hw/ncdevmem.c @@ -76,7 +76,7 @@
#define PAGE_SHIFT 12 #define TEST_PREFIX "ncdevmem" -#define NUM_PAGES 16000 +#define NUM_PAGES 64
#ifndef MSG_SOCK_DEVMEM #define MSG_SOCK_DEVMEM 0x2000000
To my surprise, the test passed just fine. Seems the limitation at least doesn't apply to GVE. I don't know what the rest of the drivers do, but so far this seems like driver specific behavior. I think putting limitations in the core stack for mlx5 issues doesn't seem great.
pool_size seems to be the number of ptr_ring slots in the page_pool, not some upper or lower bound on the amount of memory the page_pool can provide. So this check seems useless? The page_pool can still not provide this amount of memory with dmabuf (if the netmems aren't being recycled fast enough) or with normal memory (under memory pressure).
I think pool_size is usually set by drivers based on their params, and it's the max size of pool->ring. The opportunistic check I added compares this demand with the supply (available chunk memory) and fails this config based on the assumption that there should be enough memory in the pool to satisfy driver needs.
Please let me know your thoughts and how to proceed.
I think there are better options here:
1. In the page_pool, warn if the dmabuf is too small for the ring size, but don't outright prevent the configuration. If the user is running on a driver that doesn't have a dmabuf size limitation, let them ignore the warning and run it.
2. In mlx5 code, somehow find out how big the dmabuf size is (it may need a new pp api), and then in mlx5 code prevent this configuration to workaround your (I think) driver-specific limitation.
3. Maybe address the general limitation in mlx5, and make it work even if it can't refill its rings? It would help this case and other memory pressure scenarios as well.
linux-kselftest-mirror@lists.linaro.org
-
Cosmin Ratiu -
Mina Almasry -
Stanislav Fomichev