drivers/nvme/host/pci.c | 171 +++++++++++++++++++++++----------------- include/linux/dmapool.h | 17 +++- mm/dmapool.c | 16 ++-- 3 files changed, 121 insertions(+), 83 deletions(-)
NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Allocating the dmapool structs on the desired NUMA node further reduces the time spent in dma_pool_alloc from 0.87% to 0.50%. Caleb Sander Mateos (2): nvme/pci: factor out nvme_init_hctx() helper nvme/pci: make PRP list DMA pools per-NUMA-node Keith Busch (1): dmapool: add NUMA affinity support drivers/nvme/host/pci.c | 171 +++++++++++++++++++++++----------------- include/linux/dmapool.h | 17 +++- mm/dmapool.c | 16 ++-- 3 files changed, 121 insertions(+), 83 deletions(-) v6: - Clarify description of when PRP list pages are allocated (Christoph) - Add Reviewed-by tags v5: - Allocate dmapool structs on desired NUMA node (Keith) - Add Reviewed-by tags v4: - Drop the numa_node < nr_node_ids check (Kanchan) - Add Reviewed-by tags v3: simplify nvme_release_prp_pools() (Keith) v2: - Initialize admin nvme_queue's nvme_prp_dma_pools (Kanchan) - Shrink nvme_dev's prp_pools array from MAX_NUMNODES to nr_node_ids (Kanchan) -- 2.45.2
As it's been impossible to get any MM feedback even with multiple pings, so I've tentatively queued it up in nvme-6.16. If anyone in MM land doesn't like this please scream now and I'll drop it ASAP.
Hi all, It seems like there is consensus on this series and all patches have multiple reviews. Would it be possible to queue it up for 6.16? The NVMe tree seems like it would make sense, though maybe the dmapool patch needs to go through the mm tree? Thanks, Caleb On Fri, Apr 25, 2025 at 7:07 PM Caleb Sander Mateos <csander@purestorage.com> wrote: > > NVMe commands with over 8 KB of discontiguous data allocate PRP list > pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. > Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool > spinlock. These device-global spinlocks are a significant source of > contention when many CPUs are submitting to the same NVMe devices. On a > workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 > NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in > _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. > > Ideally, the dma_pools would be per-hctx to minimize contention. But > that could impose considerable resource costs in a system with many NVMe > devices and CPUs. > > As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each > nvme_queue to the set of DMA pools corresponding to its device and its > hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by > about half, to 1.2%. Preventing the sharing of PRP list pages across > NUMA nodes also makes them cheaper to initialize. > > Allocating the dmapool structs on the desired NUMA node further reduces > the time spent in dma_pool_alloc from 0.87% to 0.50%. > > Caleb Sander Mateos (2): > nvme/pci: factor out nvme_init_hctx() helper > nvme/pci: make PRP list DMA pools per-NUMA-node > > Keith Busch (1): > dmapool: add NUMA affinity support > > drivers/nvme/host/pci.c | 171 +++++++++++++++++++++++----------------- > include/linux/dmapool.h | 17 +++- > mm/dmapool.c | 16 ++-- > 3 files changed, 121 insertions(+), 83 deletions(-) > > v6: > - Clarify description of when PRP list pages are allocated (Christoph) > - Add Reviewed-by tags > > v5: > - Allocate dmapool structs on desired NUMA node (Keith) > - Add Reviewed-by tags > > v4: > - Drop the numa_node < nr_node_ids check (Kanchan) > - Add Reviewed-by tags > > v3: simplify nvme_release_prp_pools() (Keith) > > v2: > - Initialize admin nvme_queue's nvme_prp_dma_pools (Kanchan) > - Shrink nvme_dev's prp_pools array from MAX_NUMNODES to nr_node_ids (Kanchan) > > -- > 2.45.2 >
On Fri, May 02, 2025 at 09:48:17AM -0700, Caleb Sander Mateos wrote: > It seems like there is consensus on this series and all patches have > multiple reviews. Would it be possible to queue it up for 6.16? The > NVMe tree seems like it would make sense, though maybe the dmapool > patch needs to go through the mm tree? The subsequent nvme patches depend on the dmapool patch, so they all need to go through the same tree. It's okay with me if nvme picks these up, but I think we need an Ack from one of the mm reviewers/maintainers.
© 2016 - 2025 Red Hat, Inc.