| Message ID | 20240624153321.29834-1-anand.a.khoje@oracle.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [v5] net/mlx5: Reclaim max 50K pages at once | expand |
On 6/24/2024 8:33 AM, Anand Khoje wrote: > --- a/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c > +++ b/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c > @@ -608,6 +608,7 @@ enum { > RELEASE_ALL_PAGES_MASK = 0x4000, > }; > > +#define MAX_RECLAIM_NPAGES -50000 Can you please explain why this is negative? There doesn't seem to be any reason mentioned in the commit message or code. At the very least it's super confusing to have a MAX be negative, and at worst it's a bug. I don't have any other context on this code besides this patch, so an explanation would be helpful.
On 6/25/24 02:11, Jesse Brandeburg wrote: > On 6/24/2024 8:33 AM, Anand Khoje wrote: > >> --- a/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c >> +++ b/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c >> @@ -608,6 +608,7 @@ enum { >> RELEASE_ALL_PAGES_MASK = 0x4000, >> }; >> >> +#define MAX_RECLAIM_NPAGES -50000 > Can you please explain why this is negative? There doesn't seem to be > any reason mentioned in the commit message or code. > > At the very least it's super confusing to have a MAX be negative, and at > worst it's a bug. I don't have any other context on this code besides > this patch, so an explanation would be helpful. > > > Hi Jesse, The way Mellanox ConnectX5 driver handles 'release of allocated pages from HCA' or 'allocation of pages to HCA', is by sending an event to the host. This event will have number of pages in it. If the number is positive, that indicates HCA is requesting that number of pages to be allocated. And if that number is negative, it is the HCA indicating that that number of pages can be reclaimed by the host. In this patch we are restricting the maximum number of pages that can be reclaimed to be 50000 (effectively this would be -50000 as it is reclaim). This limit is based on the capability of the firmware as it cannot release more than 50000 back to the host in one go. I hope that explains. Thanks, Anand
在 2024/6/25 13:00, Anand Khoje 写道: > > On 6/25/24 02:11, Jesse Brandeburg wrote: >> On 6/24/2024 8:33 AM, Anand Khoje wrote: >> >>> --- a/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c >>> +++ b/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c >>> @@ -608,6 +608,7 @@ enum { >>> RELEASE_ALL_PAGES_MASK = 0x4000, >>> }; >>> +#define MAX_RECLAIM_NPAGES -50000 >> Can you please explain why this is negative? There doesn't seem to be >> any reason mentioned in the commit message or code. >> >> At the very least it's super confusing to have a MAX be negative, and at >> worst it's a bug. I don't have any other context on this code besides >> this patch, so an explanation would be helpful. >> >> >> > Hi Jesse, > > The way Mellanox ConnectX5 driver handles 'release of allocated pages > from HCA' or 'allocation of pages to HCA', is by sending an event to the > host. This event will have number of pages in it. If the number is > positive, that indicates HCA is requesting that number of pages to be > allocated. And if that number is negative, it is the HCA indicating that > that number of pages can be reclaimed by the host. > > In this patch we are restricting the maximum number of pages that can be > reclaimed to be 50000 (effectively this would be -50000 as it is > reclaim). This limit is based on the capability of the firmware as it > cannot release more than 50000 back to the host in one go. > > I hope that explains. To be honest, I am also obvious why this MACRO is defined as a negative number. From the above, I can understand why. I think, perhaps many people also wonder why it is defined as a negative. IMO, it is better that you put the above explanations into the source code as comments. When users check the source code, from the comments, users will know why it is defined as a negative number. Thanks a lot. Zhu Yanjun > > Thanks, > > Anand >
On Wed, Jun 26, 2024 at 04:19:17AM +0800, Zhu Yanjun wrote: > 在 2024/6/25 13:00, Anand Khoje 写道: > > > > On 6/25/24 02:11, Jesse Brandeburg wrote: > > > On 6/24/2024 8:33 AM, Anand Khoje wrote: > > > > > > > --- a/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c > > > > +++ b/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c > > > > @@ -608,6 +608,7 @@ enum { > > > > RELEASE_ALL_PAGES_MASK = 0x4000, > > > > }; > > > > +#define MAX_RECLAIM_NPAGES -50000 > > > Can you please explain why this is negative? There doesn't seem to be > > > any reason mentioned in the commit message or code. > > > > > > At the very least it's super confusing to have a MAX be negative, and at > > > worst it's a bug. I don't have any other context on this code besides > > > this patch, so an explanation would be helpful. > > > > > > > > > > > Hi Jesse, > > > > The way Mellanox ConnectX5 driver handles 'release of allocated pages > > from HCA' or 'allocation of pages to HCA', is by sending an event to the > > host. This event will have number of pages in it. If the number is > > positive, that indicates HCA is requesting that number of pages to be > > allocated. And if that number is negative, it is the HCA indicating that > > that number of pages can be reclaimed by the host. > > > > In this patch we are restricting the maximum number of pages that can be > > reclaimed to be 50000 (effectively this would be -50000 as it is > > reclaim). This limit is based on the capability of the firmware as it > > cannot release more than 50000 back to the host in one go. > > > > I hope that explains. > > To be honest, I am also obvious why this MACRO is defined as a negative > number. From the above, I can understand why. I think, perhaps many people > also wonder why it is defined as a negative. IMO, it is better that you put > the above explanations into the source code as comments. > When users check the source code, from the comments, users will know why it > is defined as a negative number. I see no problem with adding a comment to the code, but I think that it won't help anyone. The whole reclaim/give page logic inside the mlx5 driver is written with the assumption that the number of pages is negative for reclaim and positive for give. Thanks > > Thanks a lot. > Zhu Yanjun > > > > > Thanks, > > > > Anand > > > >
... > The way Mellanox ConnectX5 driver handles 'release of allocated pages > from HCA' or 'allocation of pages to HCA', is by sending an event to the > host. This event will have number of pages in it. If the number is > positive, that indicates HCA is requesting that number of pages to be > allocated. And if that number is negative, it is the HCA indicating that > that number of pages can be reclaimed by the host. A one line comment would do. Possibly even negating the be32toh() result? > In this patch we are restricting the maximum number of pages that can be > reclaimed to be 50000 (effectively this would be -50000 as it is > reclaim). This limit is based on the capability of the firmware as it > cannot release more than 50000 back to the host in one go. Hang on, why are you soft limiting it to the hard limit? I thought the problem was that releasing a lot of pages took a long time and 'stuffed' other time-critical tasks. The only way to resolve that would seem to be to defer the actual freeing to a low (or at least normal user) priority thread. You would definitely want to get out of 'softint' context. (Which is out of napi unless forced to be threaded - and that only really works if you force the threads under the RT scheduler.) David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
On 6/28/24 21:14, David Laight wrote: > ... >> The way Mellanox ConnectX5 driver handles 'release of allocated pages >> from HCA' or 'allocation of pages to HCA', is by sending an event to the >> host. This event will have number of pages in it. If the number is >> positive, that indicates HCA is requesting that number of pages to be >> allocated. And if that number is negative, it is the HCA indicating that >> that number of pages can be reclaimed by the host. > A one line comment would do. > Possibly even negating the be32toh() result? > >> In this patch we are restricting the maximum number of pages that can be >> reclaimed to be 50000 (effectively this would be -50000 as it is >> reclaim). This limit is based on the capability of the firmware as it >> cannot release more than 50000 back to the host in one go. > Hang on, why are you soft limiting it to the hard limit? > I thought the problem was that releasing a lot of pages took a long > time and 'stuffed' other time-critical tasks. > > The only way to resolve that would seem to be to defer the actual freeing > to a low (or at least normal user) priority thread. > You would definitely want to get out of 'softint' context. > (Which is out of napi unless forced to be threaded - and that only really > works if you force the threads under the RT scheduler.) > > David Hi David, The issue here is, when Mellanox device sends a huge number of pages back to the host to reclaim, the host allocates a certain number of mailbox messages mlx5_cmd_mailbox to accommodate the DMA addresses of the memory to be reclaimed. The freeing of these mailbox messages is time consuming (not the freeing of actual pages). Now, the limit of the FW is that presently, it frees upto 50000 pages. This limit can increase in future firmware versions. We are limiting this in the driver because we see optimal results with this limit during our tests. The results indicated that the time consumed while freeing of mailbox messages stayed 2 usec on average - which is tolerable and would not need running this thread in a different (low priority) context. Thanks, Anand > - > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > Registration No: 1397386 (Wales)
On 6/28/24 21:14, David Laight wrote: > ... >> The way Mellanox ConnectX5 driver handles 'release of allocated pages >> from HCA' or 'allocation of pages to HCA', is by sending an event to the >> host. This event will have number of pages in it. If the number is >> positive, that indicates HCA is requesting that number of pages to be >> allocated. And if that number is negative, it is the HCA indicating that >> that number of pages can be reclaimed by the host. > A one line comment would do. > Possibly even negating the be32toh() result? > >> In this patch we are restricting the maximum number of pages that can be >> reclaimed to be 50000 (effectively this would be -50000 as it is >> reclaim). This limit is based on the capability of the firmware as it >> cannot release more than 50000 back to the host in one go. > Hang on, why are you soft limiting it to the hard limit? > I thought the problem was that releasing a lot of pages took a long > time and 'stuffed' other time-critical tasks. > > The only way to resolve that would seem to be to defer the actual freeing > to a low (or at least normal user) priority thread. > You would definitely want to get out of 'softint' context. > (Which is out of napi unless forced to be threaded - and that only really > works if you force the threads under the RT scheduler.) > > David Hi David, The issue here is, when Mellanox device sends a huge number of pages back to the host to reclaim, the host allocates a certain number of mailbox messages mlx5_cmd_mailbox to accommodate the DMA addresses of the memory to be reclaimed. The freeing of these mailbox messages is time consuming (not the freeing of actual pages). Now, the limit of the FW is that presently, it frees upto 50000 pages. This limit can increase in future firmware versions. We are limiting this in the driver because we see optimal results with this limit during our tests. The results indicated that the time consumed while freeing of mailbox messages stayed 2 usec on average - which is tolerable and would not need running this thread in a different (low priority) context. Thanks, Anand > - > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > Registration No: 1397386 (Wales)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c b/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c index d894a88..1fc583b 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c @@ -608,6 +608,7 @@ enum { RELEASE_ALL_PAGES_MASK = 0x4000, }; +#define MAX_RECLAIM_NPAGES -50000 static int req_pages_handler(struct notifier_block *nb, unsigned long type, void *data) { @@ -639,7 +640,7 @@ static int req_pages_handler(struct notifier_block *nb, req->dev = dev; req->func_id = func_id; - req->npages = npages; + req->npages = max_t(s32, npages, MAX_RECLAIM_NPAGES); req->ec_function = ec_function; req->release_all = release_all; INIT_WORK(&req->work, pages_work_handler);
In non FLR context, at times CX-5 requests release of ~8 million FW pages. This needs humongous number of cmd mailboxes, which to be released once the pages are reclaimed. Release of humongous number of cmd mailboxes is consuming cpu time running into many seconds. Which with non preemptible kernels is leading to critical process starving on that cpu’s RQ. To alleviate this, this change restricts the total number of pages a worker will try to reclaim maximum 50K pages in one go. The limit 50K is aligned with the current firmware capacity/limit of releasing 50K pages at once per MLX5_CMD_OP_MANAGE_PAGES + MLX5_PAGES_TAKE device command. Our tests have shown significant benefit of this change in terms of time consumed by dma_pool_free(). During a test where an event was raised by HCA to release 1.3 Million pages, following observations were made: - Without this change: Number of mailbox messages allocated was around 20K, to accommodate the DMA addresses of 1.3 million pages. The average time spent by dma_pool_free() to free the DMA pool is between 16 usec to 32 usec. value ------------- Distribution ------------- count 256 | 0 512 |@ 287 1024 |@@@ 1332 2048 |@ 656 4096 |@@@@@ 2599 8192 |@@@@@@@@@@ 4755 16384 |@@@@@@@@@@@@@@@ 7545 32768 |@@@@@ 2501 65536 | 0 - With this change: Number of mailbox messages allocated was around 800; this was to accommodate DMA addresses of only 50K pages. The average time spent by dma_pool_free() to free the DMA pool in this case lies between 1 usec to 2 usec. value ------------- Distribution ------------- count 256 | 0 512 |@@@@@@@@@@@@@@@@@@ 346 1024 |@@@@@@@@@@@@@@@@@@@@@@ 435 2048 | 0 4096 | 0 8192 | 1 16384 | 0 Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com> --- Changes in v5: - Made changes as per a suggestion from Leon. --- drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)