| Message ID | 1557079171-19329-1-git-send-email-galpress@amazon.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | RDMA/efa: Elastic Fabric Adapter (EFA) driver | expand |
On Sun, May 05, 2019 at 08:59:20PM +0300, Gal Pressman wrote: > Hello all, > The following v7 patchset introduces the Amazon Elastic Fabric Adapter (EFA) > driver. > > EFA is a networking adapter designed to support user space network > communication, initially offered in the Amazon EC2 environment. First release > of EFA supports datagram send/receive operations and does not support > connection-oriented or read/write operations. > > EFA supports Unreliable Datagrams (UD) as well as a new unordered, Scalable > Reliable Datagram protocol (SRD). SRD provides support for reliable datagrams > and more complete error handling than typical RD, but, unlike RD, it does not > support ordering nor segmentation. > > EFA reliable datagram transport provides reliable out-of-order delivery, > transparently utilizing multiple network paths to reduce network tail > latency. Its interface is similar to UD, in particular it supports > message size up to MTU, with error handling extended to support reliable > communication. More information regarding SRD can be found at [1]. > > Kernel verbs and in-kernel services are initially not supported but are planned > for future releases. > > EFA enabled EC2 instances have two different devices allocated, one for ENA > (netdev) and one for EFA, the two are separate pci devices with no in-kernel > communication between them. > > This patchset also introduces RDMA subsystem ibdev_* print helpers which should > be used by the other new drivers that are currently under review (irdma, siw) > and over time by all drivers in the subsystem. > The print format is similar to the netdev_* helpers. > > PR for rdma-core provider was sent: > https://github.com/linux-rdma/rdma-core/pull/475 > > Thanks to everyone who took the time to review our last submissions (Jason, Doug, > Sean, Dennis, Leon, Christoph, Parav, Sagi, Steve, Shiraz), it is very > appreciated. > > Issues addressed in v7: > * Remove unused list field in mmap entry (Jason) > * Protect mmap xarray operations with a lock (Jason) > * Outstanding comments will be fixed in follow-up submission > > Issues addressed in v6: > * Remove BUG_ONs from __dynamic_ibdev_dbg > * Remove redundant udata checks (Leon) > * Remove rdma_user_mmap_page usage > * Remove umem->page_shift usage > * Make efa_destroy_qp_handle function static > * Misc prints cleanups > > Issues addressed in v5: > * Adapt to subsystem verbs API changes > * Remove unnecessary 'do ... while' in ibdev_dbg (Jason) > * Use a non-macro implementation for ibdev_dbg (Jason) > * Use for_each_sg_dma_page() in umem iterations (Jason) > * Remove unused enum value EFA_ADMIN_START_CMD_RANGE (Leon) > * Don't assume the sg element offset is zero (Jason) > * Cherry-picked Shiraz's new ib_umem_find_single_pg_size() work, encountered > some issues, debugging with Shiraz off-list. Will convert to use his work once > solved. > > Issues addressed in v4: > * Add RDMA subsystem ibdev_* printk helpers (Leon, Jason) > * Use xarray for mmap keys (Jason) > * Use module_pci_driver macro (Jason) > * Remove redundant cast in efa_remove (Jason) > * Avoid unnecessary use of pci_get_drvdata (Jason) > * Remove unnecessary admin queue sizes macros (Leon) > * Remove EFA_DEVICE_RUNNING bit (Leon, Jason) > * Remove incorrect comment in efa_com_validate_version (Leon) > * Keep lists sorted (Jason) > * Keep efa_com_dev as part of efa_dev instead of allocating it (Jason) > > Issues addressed in v3: > * Use new rdma_udata_to_drv_context API > * Adapt to new core ucontext allocations > * Remove EFA transport/protocol/node type and use unspecified instead (Leon, Jason) > * Replace stats lock with atomic variables (Leon, Jason, Steve) > * Remove vertical alignment from structs (Steve) > * Remove license text from ABI file (Leon) > * Undefine macro when it's no longer used (Steve) > * Fix kdoc formatting (Steve) > * Remove unneeded lock from reg read destroy flow (Steve) > * Prefer {} initializations over memset (Leon) > * Remove highmem WARN_ON_ONCE (Steve) > * ib_alloc_device returns NULL in case of error (Leon) > * Remove redundant check from remove remove device flow (Leon) > * Remove redundant zero assignments after memsets > * Remove unnecessary WARN_ON_ONCEs from create QP verbs (Steve) > * Remove redundant memsets (Steve, Shiraz) > * Change all non-privileged flows error prints to debug level (Steve, Leon, Jason, Shiraz) > * Remove likely/unlikelys from control path (Leon, Jason) > * Fixes to reg MR indirect flow wrong PAGE_SIZE usage (Jason) > * Use decimal array size in ABI file (Steve) > * Remove redundant comments (Steve, Shiraz) > * Change efa_verbs.c to GPL-2.0 OR Linux-OpenIB license (Leon, Jason) > * Replace WARN in admin completion processing with WARN_ONCE (Steve) > > Major issues addressed in this v2: > * Userspace libibverbs provider is implemented and attached for review. > * Respect the atomic requirement of create/destroy AH flows using the new > sleepable flag [2]. > * Change link layer from Ethernet to Unspecified (Proprietary EC2 link layer). > * Use RDMA mmap API. > * Coherent DMA memory is no longer mapped to the userspace, streaming DMA > mappings are used instead. > * Introduce alloc/dealloc PD admin commands, PDs are now backed by an object on > the device. This removes the bitmap used for PD number allocations. > * Addressed the mmap lifetime issues: > Each ucontext now uses a new User Access Region (UAR) abstraction. > Objects which are tied to a specific UAR will not be allocated to a different > user until the UAR is deallocated (on application exit). > DMA memory will be unmapped when the QP/CQ is destroyed, but the buffers will > remain allocated until application exit. > The mmap entries now remain valid until application exit and allow for reuse > of the same mmap key more than once. > * SRD QP type is now a driver QP type (previously was IB_QPT_SRD). > * Match UD QP Infiniband semantics, including 40 bytes offset, state transitions, > QKey validation, etc. > * Move AH reference counts to the device (previously was in the driver). > When creating more than one AH with the same GID, the same device resource is > used internally. Instead of keeping the reference count in the driver (and issue > one create AH command only), each AH creation is now passed on to the device > (accompanied with the PD number). > This allows for future optimizations for AHs that are no longer used by a > specific PD. > * Removed all stub functions, which will mark EFA driver as a non-kverbs provider [3]. > * Replace all pr_* prints with dev_* prints > > [1] https://github.com/amzn/rdma-core/wiki/SRD > [2] https://patchwork.kernel.org/cover/10725727/ > [3] https://patchwork.kernel.org/cover/10775039/ > > Thanks, > Gal > > Gal Pressman (11): > RDMA: Add EFA related definitions > RDMA/efa: Add EFA device definitions > RDMA/efa: Add the efa.h header file > RDMA/efa: Add the efa_com.h file > RDMA/efa: Add the com service API definitions > RDMA/efa: Add the ABI definitions > RDMA/efa: Implement functions that submit and complete admin commands > RDMA/efa: Add common command handlers > RDMA/efa: Add EFA verbs implementation > RDMA/efa: Add the efa module > RDMA/efa: Add driver to Kconfig/Makefile I've tentatively applied this to for-next, pending your ack of my changes. Thanks, Jason
On 06-May-19 21:20, Jason Gunthorpe wrote: > On Sun, May 05, 2019 at 08:59:20PM +0300, Gal Pressman wrote: >> Hello all, >> The following v7 patchset introduces the Amazon Elastic Fabric Adapter (EFA) >> driver. >> >> EFA is a networking adapter designed to support user space network >> communication, initially offered in the Amazon EC2 environment. First release >> of EFA supports datagram send/receive operations and does not support >> connection-oriented or read/write operations. >> >> EFA supports Unreliable Datagrams (UD) as well as a new unordered, Scalable >> Reliable Datagram protocol (SRD). SRD provides support for reliable datagrams >> and more complete error handling than typical RD, but, unlike RD, it does not >> support ordering nor segmentation. >> >> EFA reliable datagram transport provides reliable out-of-order delivery, >> transparently utilizing multiple network paths to reduce network tail >> latency. Its interface is similar to UD, in particular it supports >> message size up to MTU, with error handling extended to support reliable >> communication. More information regarding SRD can be found at [1]. >> >> Kernel verbs and in-kernel services are initially not supported but are planned >> for future releases. >> >> EFA enabled EC2 instances have two different devices allocated, one for ENA >> (netdev) and one for EFA, the two are separate pci devices with no in-kernel >> communication between them. >> >> This patchset also introduces RDMA subsystem ibdev_* print helpers which should >> be used by the other new drivers that are currently under review (irdma, siw) >> and over time by all drivers in the subsystem. >> The print format is similar to the netdev_* helpers. >> >> PR for rdma-core provider was sent: >> https://github.com/linux-rdma/rdma-core/pull/475 >> >> Thanks to everyone who took the time to review our last submissions (Jason, Doug, >> Sean, Dennis, Leon, Christoph, Parav, Sagi, Steve, Shiraz), it is very >> appreciated. >> >> Issues addressed in v7: >> * Remove unused list field in mmap entry (Jason) >> * Protect mmap xarray operations with a lock (Jason) >> * Outstanding comments will be fixed in follow-up submission >> >> Issues addressed in v6: >> * Remove BUG_ONs from __dynamic_ibdev_dbg >> * Remove redundant udata checks (Leon) >> * Remove rdma_user_mmap_page usage >> * Remove umem->page_shift usage >> * Make efa_destroy_qp_handle function static >> * Misc prints cleanups >> >> Issues addressed in v5: >> * Adapt to subsystem verbs API changes >> * Remove unnecessary 'do ... while' in ibdev_dbg (Jason) >> * Use a non-macro implementation for ibdev_dbg (Jason) >> * Use for_each_sg_dma_page() in umem iterations (Jason) >> * Remove unused enum value EFA_ADMIN_START_CMD_RANGE (Leon) >> * Don't assume the sg element offset is zero (Jason) >> * Cherry-picked Shiraz's new ib_umem_find_single_pg_size() work, encountered >> some issues, debugging with Shiraz off-list. Will convert to use his work once >> solved. >> >> Issues addressed in v4: >> * Add RDMA subsystem ibdev_* printk helpers (Leon, Jason) >> * Use xarray for mmap keys (Jason) >> * Use module_pci_driver macro (Jason) >> * Remove redundant cast in efa_remove (Jason) >> * Avoid unnecessary use of pci_get_drvdata (Jason) >> * Remove unnecessary admin queue sizes macros (Leon) >> * Remove EFA_DEVICE_RUNNING bit (Leon, Jason) >> * Remove incorrect comment in efa_com_validate_version (Leon) >> * Keep lists sorted (Jason) >> * Keep efa_com_dev as part of efa_dev instead of allocating it (Jason) >> >> Issues addressed in v3: >> * Use new rdma_udata_to_drv_context API >> * Adapt to new core ucontext allocations >> * Remove EFA transport/protocol/node type and use unspecified instead (Leon, Jason) >> * Replace stats lock with atomic variables (Leon, Jason, Steve) >> * Remove vertical alignment from structs (Steve) >> * Remove license text from ABI file (Leon) >> * Undefine macro when it's no longer used (Steve) >> * Fix kdoc formatting (Steve) >> * Remove unneeded lock from reg read destroy flow (Steve) >> * Prefer {} initializations over memset (Leon) >> * Remove highmem WARN_ON_ONCE (Steve) >> * ib_alloc_device returns NULL in case of error (Leon) >> * Remove redundant check from remove remove device flow (Leon) >> * Remove redundant zero assignments after memsets >> * Remove unnecessary WARN_ON_ONCEs from create QP verbs (Steve) >> * Remove redundant memsets (Steve, Shiraz) >> * Change all non-privileged flows error prints to debug level (Steve, Leon, Jason, Shiraz) >> * Remove likely/unlikelys from control path (Leon, Jason) >> * Fixes to reg MR indirect flow wrong PAGE_SIZE usage (Jason) >> * Use decimal array size in ABI file (Steve) >> * Remove redundant comments (Steve, Shiraz) >> * Change efa_verbs.c to GPL-2.0 OR Linux-OpenIB license (Leon, Jason) >> * Replace WARN in admin completion processing with WARN_ONCE (Steve) >> >> Major issues addressed in this v2: >> * Userspace libibverbs provider is implemented and attached for review. >> * Respect the atomic requirement of create/destroy AH flows using the new >> sleepable flag [2]. >> * Change link layer from Ethernet to Unspecified (Proprietary EC2 link layer). >> * Use RDMA mmap API. >> * Coherent DMA memory is no longer mapped to the userspace, streaming DMA >> mappings are used instead. >> * Introduce alloc/dealloc PD admin commands, PDs are now backed by an object on >> the device. This removes the bitmap used for PD number allocations. >> * Addressed the mmap lifetime issues: >> Each ucontext now uses a new User Access Region (UAR) abstraction. >> Objects which are tied to a specific UAR will not be allocated to a different >> user until the UAR is deallocated (on application exit). >> DMA memory will be unmapped when the QP/CQ is destroyed, but the buffers will >> remain allocated until application exit. >> The mmap entries now remain valid until application exit and allow for reuse >> of the same mmap key more than once. >> * SRD QP type is now a driver QP type (previously was IB_QPT_SRD). >> * Match UD QP Infiniband semantics, including 40 bytes offset, state transitions, >> QKey validation, etc. >> * Move AH reference counts to the device (previously was in the driver). >> When creating more than one AH with the same GID, the same device resource is >> used internally. Instead of keeping the reference count in the driver (and issue >> one create AH command only), each AH creation is now passed on to the device >> (accompanied with the PD number). >> This allows for future optimizations for AHs that are no longer used by a >> specific PD. >> * Removed all stub functions, which will mark EFA driver as a non-kverbs provider [3]. >> * Replace all pr_* prints with dev_* prints >> >> [1] https://github.com/amzn/rdma-core/wiki/SRD >> [2] https://patchwork.kernel.org/cover/10725727/ >> [3] https://patchwork.kernel.org/cover/10775039/ >> >> Thanks, >> Gal >> >> Gal Pressman (11): >> RDMA: Add EFA related definitions >> RDMA/efa: Add EFA device definitions >> RDMA/efa: Add the efa.h header file >> RDMA/efa: Add the efa_com.h file >> RDMA/efa: Add the com service API definitions >> RDMA/efa: Add the ABI definitions >> RDMA/efa: Implement functions that submit and complete admin commands >> RDMA/efa: Add common command handlers >> RDMA/efa: Add EFA verbs implementation >> RDMA/efa: Add the efa module >> RDMA/efa: Add driver to Kconfig/Makefile > > I've tentatively applied this to for-next, pending your ack of my > changes. Thanks! Other than the minor fixes I posted the changes looks good. Acked-by: Gal Pressman <galpress@amazon.com>
Hello all, The following v7 patchset introduces the Amazon Elastic Fabric Adapter (EFA) driver. EFA is a networking adapter designed to support user space network communication, initially offered in the Amazon EC2 environment. First release of EFA supports datagram send/receive operations and does not support connection-oriented or read/write operations. EFA supports Unreliable Datagrams (UD) as well as a new unordered, Scalable Reliable Datagram protocol (SRD). SRD provides support for reliable datagrams and more complete error handling than typical RD, but, unlike RD, it does not support ordering nor segmentation. EFA reliable datagram transport provides reliable out-of-order delivery, transparently utilizing multiple network paths to reduce network tail latency. Its interface is similar to UD, in particular it supports message size up to MTU, with error handling extended to support reliable communication. More information regarding SRD can be found at [1]. Kernel verbs and in-kernel services are initially not supported but are planned for future releases. EFA enabled EC2 instances have two different devices allocated, one for ENA (netdev) and one for EFA, the two are separate pci devices with no in-kernel communication between them. This patchset also introduces RDMA subsystem ibdev_* print helpers which should be used by the other new drivers that are currently under review (irdma, siw) and over time by all drivers in the subsystem. The print format is similar to the netdev_* helpers. PR for rdma-core provider was sent: https://github.com/linux-rdma/rdma-core/pull/475 Thanks to everyone who took the time to review our last submissions (Jason, Doug, Sean, Dennis, Leon, Christoph, Parav, Sagi, Steve, Shiraz), it is very appreciated. Issues addressed in v7: * Remove unused list field in mmap entry (Jason) * Protect mmap xarray operations with a lock (Jason) * Outstanding comments will be fixed in follow-up submission Issues addressed in v6: * Remove BUG_ONs from __dynamic_ibdev_dbg * Remove redundant udata checks (Leon) * Remove rdma_user_mmap_page usage * Remove umem->page_shift usage * Make efa_destroy_qp_handle function static * Misc prints cleanups Issues addressed in v5: * Adapt to subsystem verbs API changes * Remove unnecessary 'do ... while' in ibdev_dbg (Jason) * Use a non-macro implementation for ibdev_dbg (Jason) * Use for_each_sg_dma_page() in umem iterations (Jason) * Remove unused enum value EFA_ADMIN_START_CMD_RANGE (Leon) * Don't assume the sg element offset is zero (Jason) * Cherry-picked Shiraz's new ib_umem_find_single_pg_size() work, encountered some issues, debugging with Shiraz off-list. Will convert to use his work once solved. Issues addressed in v4: * Add RDMA subsystem ibdev_* printk helpers (Leon, Jason) * Use xarray for mmap keys (Jason) * Use module_pci_driver macro (Jason) * Remove redundant cast in efa_remove (Jason) * Avoid unnecessary use of pci_get_drvdata (Jason) * Remove unnecessary admin queue sizes macros (Leon) * Remove EFA_DEVICE_RUNNING bit (Leon, Jason) * Remove incorrect comment in efa_com_validate_version (Leon) * Keep lists sorted (Jason) * Keep efa_com_dev as part of efa_dev instead of allocating it (Jason) Issues addressed in v3: * Use new rdma_udata_to_drv_context API * Adapt to new core ucontext allocations * Remove EFA transport/protocol/node type and use unspecified instead (Leon, Jason) * Replace stats lock with atomic variables (Leon, Jason, Steve) * Remove vertical alignment from structs (Steve) * Remove license text from ABI file (Leon) * Undefine macro when it's no longer used (Steve) * Fix kdoc formatting (Steve) * Remove unneeded lock from reg read destroy flow (Steve) * Prefer {} initializations over memset (Leon) * Remove highmem WARN_ON_ONCE (Steve) * ib_alloc_device returns NULL in case of error (Leon) * Remove redundant check from remove remove device flow (Leon) * Remove redundant zero assignments after memsets * Remove unnecessary WARN_ON_ONCEs from create QP verbs (Steve) * Remove redundant memsets (Steve, Shiraz) * Change all non-privileged flows error prints to debug level (Steve, Leon, Jason, Shiraz) * Remove likely/unlikelys from control path (Leon, Jason) * Fixes to reg MR indirect flow wrong PAGE_SIZE usage (Jason) * Use decimal array size in ABI file (Steve) * Remove redundant comments (Steve, Shiraz) * Change efa_verbs.c to GPL-2.0 OR Linux-OpenIB license (Leon, Jason) * Replace WARN in admin completion processing with WARN_ONCE (Steve) Major issues addressed in this v2: * Userspace libibverbs provider is implemented and attached for review. * Respect the atomic requirement of create/destroy AH flows using the new sleepable flag [2]. * Change link layer from Ethernet to Unspecified (Proprietary EC2 link layer). * Use RDMA mmap API. * Coherent DMA memory is no longer mapped to the userspace, streaming DMA mappings are used instead. * Introduce alloc/dealloc PD admin commands, PDs are now backed by an object on the device. This removes the bitmap used for PD number allocations. * Addressed the mmap lifetime issues: Each ucontext now uses a new User Access Region (UAR) abstraction. Objects which are tied to a specific UAR will not be allocated to a different user until the UAR is deallocated (on application exit). DMA memory will be unmapped when the QP/CQ is destroyed, but the buffers will remain allocated until application exit. The mmap entries now remain valid until application exit and allow for reuse of the same mmap key more than once. * SRD QP type is now a driver QP type (previously was IB_QPT_SRD). * Match UD QP Infiniband semantics, including 40 bytes offset, state transitions, QKey validation, etc. * Move AH reference counts to the device (previously was in the driver). When creating more than one AH with the same GID, the same device resource is used internally. Instead of keeping the reference count in the driver (and issue one create AH command only), each AH creation is now passed on to the device (accompanied with the PD number). This allows for future optimizations for AHs that are no longer used by a specific PD. * Removed all stub functions, which will mark EFA driver as a non-kverbs provider [3]. * Replace all pr_* prints with dev_* prints [1] https://github.com/amzn/rdma-core/wiki/SRD [2] https://patchwork.kernel.org/cover/10725727/ [3] https://patchwork.kernel.org/cover/10775039/ Thanks, Gal Gal Pressman (11): RDMA: Add EFA related definitions RDMA/efa: Add EFA device definitions RDMA/efa: Add the efa.h header file RDMA/efa: Add the efa_com.h file RDMA/efa: Add the com service API definitions RDMA/efa: Add the ABI definitions RDMA/efa: Implement functions that submit and complete admin commands RDMA/efa: Add common command handlers RDMA/efa: Add EFA verbs implementation RDMA/efa: Add the efa module RDMA/efa: Add driver to Kconfig/Makefile MAINTAINERS | 9 + drivers/infiniband/Kconfig | 1 + drivers/infiniband/core/sysfs.c | 1 + drivers/infiniband/core/verbs.c | 2 + drivers/infiniband/hw/Makefile | 1 + drivers/infiniband/hw/efa/Kconfig | 15 + drivers/infiniband/hw/efa/Makefile | 9 + drivers/infiniband/hw/efa/efa.h | 162 ++ drivers/infiniband/hw/efa/efa_admin_cmds_defs.h | 794 ++++++++++ drivers/infiniband/hw/efa/efa_admin_defs.h | 136 ++ drivers/infiniband/hw/efa/efa_com.c | 1160 ++++++++++++++ drivers/infiniband/hw/efa/efa_com.h | 144 ++ drivers/infiniband/hw/efa/efa_com_cmd.c | 692 +++++++++ drivers/infiniband/hw/efa/efa_com_cmd.h | 270 ++++ drivers/infiniband/hw/efa/efa_common_defs.h | 18 + drivers/infiniband/hw/efa/efa_main.c | 533 +++++++ drivers/infiniband/hw/efa/efa_regs_defs.h | 113 ++ drivers/infiniband/hw/efa/efa_verbs.c | 1871 +++++++++++++++++++++++ include/rdma/ib_verbs.h | 4 +- include/uapi/rdma/efa-abi.h | 101 ++ include/uapi/rdma/rdma_user_ioctl_cmds.h | 1 + 21 files changed, 6036 insertions(+), 1 deletion(-) create mode 100644 drivers/infiniband/hw/efa/Kconfig create mode 100644 drivers/infiniband/hw/efa/Makefile create mode 100644 drivers/infiniband/hw/efa/efa.h create mode 100644 drivers/infiniband/hw/efa/efa_admin_cmds_defs.h create mode 100644 drivers/infiniband/hw/efa/efa_admin_defs.h create mode 100644 drivers/infiniband/hw/efa/efa_com.c create mode 100644 drivers/infiniband/hw/efa/efa_com.h create mode 100644 drivers/infiniband/hw/efa/efa_com_cmd.c create mode 100644 drivers/infiniband/hw/efa/efa_com_cmd.h create mode 100644 drivers/infiniband/hw/efa/efa_common_defs.h create mode 100644 drivers/infiniband/hw/efa/efa_main.c create mode 100644 drivers/infiniband/hw/efa/efa_regs_defs.h create mode 100644 drivers/infiniband/hw/efa/efa_verbs.c create mode 100644 include/uapi/rdma/efa-abi.h