- Kernel - mailweb.openeuler.org

[PATCH openEuler-22.03-LTS 0/9] scsi: mpt3sas: Driver patch set for openEuler-22.03-LTS
by Hao Zhang 25 Oct '23

25 Oct '23

Get patch set from upstream 5.10 LTS stable branch Ranjan Kumar (1): scsi: mpt3sas: Perform additional retries if doorbell read returns 0 Sreekanth Reddy (4): scsi: mpt3sas: Fix use-after-free warning scsi: mpt3sas: Don't change DMA mask while reallocating pools scsi: mpt3sas: re-do lost mpt3sas DMA mask fix scsi: mpt3sas: Remove usage of dma_get_required_mask() API Suganath Prabu S (1): scsi: mpt3sas: Force PCIe scatterlist allocations to be within same 4 GB region Tomas Henzl (1): scsi: mpt3sas: Fix a memory leak Wenchao Hao (1): scsi: mpt3sas: Fix NULL pointer access in mpt3sas_transport_port_add() Yang Yingliang (1): scsi: mpt3sas: Fix possible resource leaks in mpt3sas_transport_port_add() drivers/scsi/mpt3sas/mpt3sas_base.c | 222 ++++++++++++++++------- drivers/scsi/mpt3sas/mpt3sas_base.h | 2 + drivers/scsi/mpt3sas/mpt3sas_scsih.c | 2 +- drivers/scsi/mpt3sas/mpt3sas_transport.c | 16 +- 4 files changed, 175 insertions(+), 67 deletions(-) -- 2.33.0

2 10

[PATCH openEuler-1.0-LTS 0/4] Avoid spin or livelock during panic
by Ye Weihua 25 Oct '23

25 Oct '23

Stephen Brennan (4): printk: Add panic_in_progress helper printk: disable optimistic spin during panic printk: Avoid livelock with heavy printk during panic printk: Drop console_sem during panic kernel/printk/printk.c | 55 +++++++++++++++++++++++++++++++++++++++++- 1 file changed, 54 insertions(+), 1 deletion(-) -- 2.34.1

2 5

[PATCH openEuler-1.0-LTS] xen/events: replace evtchn_rwlock with RCU
by Liu Shixin 25 Oct '23

25 Oct '23

From: Juergen Gross <jgross(a)suse.com> stable inclusion from stable-v4.19.296 commit 3fdf2be9089b5096a28e76376656c60ce410ac4a category: bugfix bugzilla: 189304, https://gitee.com/src-openeuler/kernel/issues/I896EC CVE: CVE-2023-34324 Reference: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id… -------------------------------- commit 87797fad6cce28ec9be3c13f031776ff4f104cfc upstream. In unprivileged Xen guests event handling can cause a deadlock with Xen console handling. The evtchn_rwlock and the hvc_lock are taken in opposite sequence in __hvc_poll() and in Xen console IRQ handling. Normally this is no problem, as the evtchn_rwlock is taken as a reader in both paths, but as soon as an event channel is being closed, the lock will be taken as a writer, which will cause read_lock() to block: CPU0 CPU1 CPU2 (IRQ handling) (__hvc_poll()) (closing event channel) read_lock(evtchn_rwlock) spin_lock(hvc_lock) write_lock(evtchn_rwlock) [blocks] spin_lock(hvc_lock) [blocks] read_lock(evtchn_rwlock) [blocks due to writer waiting, and not in_interrupt()] This issue can be avoided by replacing evtchn_rwlock with RCU in xen_free_irq(). Note that RCU is used only to delay freeing of the irq_info memory. There is no RCU based dereferencing or replacement of pointers involved. In order to avoid potential races between removing the irq_info reference and handling of interrupts, set the irq_info pointer to NULL only when freeing its memory. The IRQ itself must be freed at that time, too, as otherwise the same IRQ number could be allocated again before handling of the old instance would have been finished. This is XSA-441 / CVE-2023-34324. Fixes: 54c9de89895e ("xen/events: add a new "late EOI" evtchn framework") Reported-by: Marek Marczykowski-Górecki <marmarek(a)invisiblethingslab.com> Signed-off-by: Juergen Gross <jgross(a)suse.com> Reviewed-by: Julien Grall <jgrall(a)amazon.com> Signed-off-by: Greg Kroah-Hartman <gregkh(a)linuxfoundation.org> Conflicts: drivers/xen/events/events_base.c drivers/xen/events/events_internal.h [ Replace set_info_for_irq() with irq_set_handler_data() ] Signed-off-by: Liu Shixin <liushixin2(a)huawei.com> --- drivers/xen/events/events_base.c | 84 ++++++++++++++-------------- drivers/xen/events/events_internal.h | 2 + 2 files changed, 45 insertions(+), 41 deletions(-) diff --git a/drivers/xen/events/events_base.c b/drivers/xen/events/events_base.c index 1cb2bf5e5e5b6..9c0ade529ec3f 100644 --- a/drivers/xen/events/events_base.c +++ b/drivers/xen/events/events_base.c @@ -82,22 +82,12 @@ const struct evtchn_ops *evtchn_ops; */ static DEFINE_MUTEX(irq_mapping_update_lock); -/* - * Lock protecting event handling loop against removing event channels. - * Adding of event channels is no issue as the associated IRQ becomes active - * only after everything is setup (before request_[threaded_]irq() the handler - * can't be entered for an event, as the event channel will be unmasked only - * then). - */ -static DEFINE_RWLOCK(evtchn_rwlock); - /* * Lock hierarchy: * * irq_mapping_update_lock - * evtchn_rwlock - * IRQ-desc lock - * percpu eoi_list_lock + * IRQ-desc lock + * percpu eoi_list_lock */ static LIST_HEAD(xen_irq_list_head); @@ -191,6 +181,22 @@ struct irq_info *info_for_irq(unsigned irq) return irq_get_handler_data(irq); } +static void delayed_free_irq(struct work_struct *work) +{ + struct irq_info *info = container_of(to_rcu_work(work), struct irq_info, + rwork); + unsigned int irq = info->irq; + + /* Remove the info pointer only now, with no potential users left. */ + irq_set_handler_data(irq, NULL); + + kfree(info); + + /* Legacy IRQ descriptors are managed by the arch. */ + if (irq >= nr_legacy_irqs()) + irq_free_desc(irq); +} + /* Constructors for packed IRQ information. */ static int xen_irq_info_common_setup(struct irq_info *info, unsigned irq, @@ -474,33 +480,36 @@ static void xen_irq_lateeoi_worker(struct work_struct *work) eoi = container_of(to_delayed_work(work), struct lateeoi_work, delayed); - read_lock_irqsave(&evtchn_rwlock, flags); + rcu_read_lock(); while (true) { - spin_lock(&eoi->eoi_list_lock); + spin_lock_irqsave(&eoi->eoi_list_lock, flags); info = list_first_entry_or_null(&eoi->eoi_list, struct irq_info, eoi_list); - if (info == NULL || now < info->eoi_time) { - spin_unlock(&eoi->eoi_list_lock); + if (info == NULL) + break; + + if (now < info->eoi_time) { + mod_delayed_work_on(info->eoi_cpu, system_wq, + &eoi->delayed, + info->eoi_time - now); break; } list_del_init(&info->eoi_list); - spin_unlock(&eoi->eoi_list_lock); + spin_unlock_irqrestore(&eoi->eoi_list_lock, flags); info->eoi_time = 0; xen_irq_lateeoi_locked(info); } - if (info) - mod_delayed_work_on(info->eoi_cpu, system_wq, - &eoi->delayed, info->eoi_time - now); + spin_unlock_irqrestore(&eoi->eoi_list_lock, flags); - read_unlock_irqrestore(&evtchn_rwlock, flags); + rcu_read_unlock(); } static void xen_cpu_init_eoi(unsigned int cpu) @@ -515,16 +524,15 @@ static void xen_cpu_init_eoi(unsigned int cpu) void xen_irq_lateeoi(unsigned int irq, unsigned int eoi_flags) { struct irq_info *info; - unsigned long flags; - read_lock_irqsave(&evtchn_rwlock, flags); + rcu_read_lock(); info = info_for_irq(irq); if (info) xen_irq_lateeoi_locked(info); - read_unlock_irqrestore(&evtchn_rwlock, flags); + rcu_read_unlock(); } EXPORT_SYMBOL_GPL(xen_irq_lateeoi); @@ -543,6 +551,7 @@ static void xen_irq_init(unsigned irq) info->type = IRQT_UNBOUND; info->refcnt = -1; + INIT_RCU_WORK(&info->rwork, delayed_free_irq); irq_set_handler_data(irq, info); @@ -595,31 +604,18 @@ static int __must_check xen_allocate_irq_gsi(unsigned gsi) static void xen_free_irq(unsigned irq) { struct irq_info *info = irq_get_handler_data(irq); - unsigned long flags; if (WARN_ON(!info)) return; - write_lock_irqsave(&evtchn_rwlock, flags); - if (!list_empty(&info->eoi_list)) lateeoi_list_del(info); list_del(&info->list); - irq_set_handler_data(irq, NULL); - WARN_ON(info->refcnt > 0); - write_unlock_irqrestore(&evtchn_rwlock, flags); - - kfree(info); - - /* Legacy IRQ descriptors are managed by the arch. */ - if (irq < nr_legacy_irqs()) - return; - - irq_free_desc(irq); + queue_rcu_work(system_wq, &info->rwork); } static void xen_evtchn_close(unsigned int port) @@ -1521,7 +1517,14 @@ static void __xen_evtchn_do_upcall(void) struct evtchn_loop_ctrl ctrl = { 0 }; unsigned count; - read_lock(&evtchn_rwlock); + /* + * When closing an event channel the associated IRQ must not be freed + * until all cpus have left the event handling loop. This is ensured + * by taking the rcu_read_lock() while handling events, as freeing of + * the IRQ is handled via queue_rcu_work() _after_ closing the event + * channel. + */ + rcu_read_lock(); do { vcpu_info->evtchn_upcall_pending = 0; @@ -1538,8 +1541,7 @@ static void __xen_evtchn_do_upcall(void) } while (count != 1 || vcpu_info->evtchn_upcall_pending); out: - - read_unlock(&evtchn_rwlock); + rcu_read_unlock(); /* * Increment irq_epoch only now to defer EOIs only for diff --git a/drivers/xen/events/events_internal.h b/drivers/xen/events/events_internal.h index 2cb9c2d2c5c0f..72e13f1638a36 100644 --- a/drivers/xen/events/events_internal.h +++ b/drivers/xen/events/events_internal.h @@ -8,6 +8,7 @@ */ #ifndef __EVENTS_INTERNAL_H__ #define __EVENTS_INTERNAL_H__ +#include <linux/rcupdate.h> /* Interrupt types. */ enum xen_irq_type { @@ -33,6 +34,7 @@ enum xen_irq_type { struct irq_info { struct list_head list; struct list_head eoi_list; + struct rcu_work rwork; int refcnt; enum xen_irq_type type; /* type */ unsigned irq; -- 2.25.1

2 1

[PATCH OLK-5.10 0/7] Synchronize coresight driver bugfix patches to openEuler
by Junhao He 25 Oct '23

25 Oct '23

1. Fix memory leak coresight: core: fix memory leak in dict->fwnode_list coresight: Fix memory leak in acpi_buffer->pointer 2. Fix TRBE bug coresight: trbe: Fix TRBE potential sleep in atomic context coresight: trbe: Allocate platform data per device coresight: trbe: Fix return value check in arm_trbe_register_coresight_cpu() 3. Fix ETR bug coresight: Fix run time warnings while reusing ETR buffer 4. acpi: Ignore the absence of graph coresight: platform: acpi: Ignore the absence of graph Revert "coresight: Return the pointer of @pdata when not "fwnode"" Junhao He (3): Revert "coresight: Return the pointer of @pdata when not "fwnode"" coresight: Fix memory leak in acpi_buffer->pointer coresight: trbe: Fix TRBE potential sleep in atomic context Linu Cherian (1): coresight: Fix run time warnings while reusing ETR buffer Suzuki K Poulose (2): coresight: trbe: Allocate platform data per device coresight: platform: acpi: Ignore the absence of graph Wei Yongjun (1): coresight: trbe: Fix return value check in arm_trbe_register_coresight_cpu() .../hwtracing/coresight/coresight-platform.c | 56 ++++++++++++------- .../hwtracing/coresight/coresight-tmc-etr.c | 20 +++---- drivers/hwtracing/coresight/coresight-trbe.c | 42 +++++++------- 3 files changed, 68 insertions(+), 50 deletions(-) -- 2.33.0

2 8

[PATCH openEuler-22.03-LTS 0/9] scsi: mpt3sas: Driver patch set for openEuler-22.03-LTS
by Hao Zhang 25 Oct '23

25 Oct '23

Get patch set from upstream 5.10 LTS stable branch Ranjan Kumar (1): scsi: mpt3sas: Perform additional retries if doorbell read returns 0 Sreekanth Reddy (4): scsi: mpt3sas: Fix use-after-free warning scsi: mpt3sas: Don't change DMA mask while reallocating pools scsi: mpt3sas: re-do lost mpt3sas DMA mask fix scsi: mpt3sas: Remove usage of dma_get_required_mask() API Suganath Prabu S (1): scsi: mpt3sas: Force PCIe scatterlist allocations to be within same 4 GB region Tomas Henzl (1): scsi: mpt3sas: Fix a memory leak Wenchao Hao (1): scsi: mpt3sas: Fix NULL pointer access in mpt3sas_transport_port_add() Yang Yingliang (1): scsi: mpt3sas: Fix possible resource leaks in mpt3sas_transport_port_add() drivers/scsi/mpt3sas/mpt3sas_base.c | 222 ++++++++++++++++------- drivers/scsi/mpt3sas/mpt3sas_base.h | 2 + drivers/scsi/mpt3sas/mpt3sas_scsih.c | 2 +- drivers/scsi/mpt3sas/mpt3sas_transport.c | 16 +- 4 files changed, 175 insertions(+), 67 deletions(-) -- 2.33.0

2 10

[PATCH OLK-5.10] perf auxtrace: Record whether an auxtrace mmap is needed
by Junhao He 25 Oct '23

25 Oct '23

From: Qi Liu <liuqi115(a)huawei.com> mainline inclusion from mainline-v6.1-rc1 commit 057381a7ece1b2726509ce47cdb9c1a111acfce9 category: Bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I7HI80 CVE: NA Reference: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id… -------------------------------- Remove ->idx() per_cpu parameter because it isn't needed. Add a flag needs_auxtrace_mmap to record whether an auxtrace mmap is needed, in preparation for correctly determining whether or not an auxtrace mmap is needed. Reviewed-by: Leo Yan <leo.yan(a)linaro.org> Signed-off-by: Qi Liu <liuqi115(a)huawei.com> Signed-off-by: Yicong Yang <yangyicong(a)hisilicon.com> Acked-by: John Garry <john.garry(a)huawei.com> Link: https://lore.kernel.org/r/20220927081400.14364-3-yangyicong@huawei.com Signed-off-by: Arnaldo Carvalho de Melo <acme(a)redhat.com> Signed-off-by: Junhao He <hejunhao3(a)huawei.com> Auxtrace mmap is needed is added to the source patch [1], However, it is deleted when upstream openEuler. Therefore, this patch adds this modification to OpenEuler OLK-5.10. [1] 057381a7ece1 "perf auxtrace arm64: Add support for HiSilicon PCIe Tune and Trace device driver" --- tools/perf/arch/arm64/util/hisi-ptt.c | 1 + 1 file changed, 1 insertion(+) diff --git a/tools/perf/arch/arm64/util/hisi-ptt.c b/tools/perf/arch/arm64/util/hisi-ptt.c index 110b2edf3e6b..97696b6e1dd5 100644 --- a/tools/perf/arch/arm64/util/hisi-ptt.c +++ b/tools/perf/arch/arm64/util/hisi-ptt.c @@ -113,6 +113,7 @@ static int hisi_ptt_recording_options(struct auxtrace_record *itr, } evsel->core.attr.freq = 0; evsel->core.attr.sample_period = 1; + evsel->needs_auxtrace_mmap = true; hisi_ptt_evsel = evsel; opts->full_auxtrace = true; } -- 2.33.0

2 1

[PATCH openEuler-1.0-LTS 0/4] Avoid spin or livelock during panic
by Ye Weihua 25 Oct '23

25 Oct '23

Stephen Brennan (4): printk: Add panic_in_progress helper printk: disable optimistic spin during panic printk: Avoid livelock with heavy printk during panic printk: Drop console_sem during panic kernel/printk/printk.c | 55 +++++++++++++++++++++++++++++++++++++++++- 1 file changed, 54 insertions(+), 1 deletion(-) -- 2.34.1

2 5

[PATCH OLK-5.10] psi: fix "no previous prototype" warnings when CONFIG_CGROUPS=n
by Lu Jialin 25 Oct '23

25 Oct '23

From: Suren Baghdasaryan <surenb(a)google.com> mainline inclusion from mainline-v5.17-rc2 commit 51e50fbd3efc6064c30ed73a5e009018b36e290a category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I89DEJ CVE: NA Reference: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?… ------------------------------------------------- When CONFIG_CGROUPS is disabled psi code generates the following warnings: kernel/sched/psi.c:1112:21: warning: no previous prototype for 'psi_trigger_create' [-Wmissing-prototypes] 1112 | struct psi_trigger *psi_trigger_create(struct psi_group *group, | ^~~~~~~~~~~~~~~~~~ kernel/sched/psi.c:1182:6: warning: no previous prototype for 'psi_trigger_destroy' [-Wmissing-prototypes] 1182 | void psi_trigger_destroy(struct psi_trigger *t) | ^~~~~~~~~~~~~~~~~~~ kernel/sched/psi.c:1249:10: warning: no previous prototype for 'psi_trigger_poll' [-Wmissing-prototypes] 1249 | __poll_t psi_trigger_poll(void **trigger_ptr, | ^~~~~~~~~~~~~~~~ Change the declarations of these functions in the header to provide the prototypes even when they are unused. Link: https://lkml.kernel.org/r/20220119223940.787748-2-surenb@google.com Fixes: 0e94682b73bf ("psi: introduce psi monitor") Signed-off-by: Suren Baghdasaryan <surenb(a)google.com> Reported-by: kernel test robot <lkp(a)intel.com> Acked-by: Johannes Weiner <hannes(a)cmpxchg.org> Signed-off-by: Andrew Morton <akpm(a)linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds(a)linux-foundation.org> Conflict: include/linux/psi.h Signed-off-by: Lu Jialin <lujialin4(a)huawei.com> --- include/linux/psi.h | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-) diff --git a/include/linux/psi.h b/include/linux/psi.h index 41079ebe17a6..86635a5630ba 100644 --- a/include/linux/psi.h +++ b/include/linux/psi.h @@ -26,12 +26,6 @@ void psi_memstall_enter(unsigned long *flags); void psi_memstall_leave(unsigned long *flags); int psi_show(struct seq_file *s, struct psi_group *group, enum psi_res res); - -#ifdef CONFIG_CGROUPS -int psi_cgroup_alloc(struct cgroup *cgrp); -void psi_cgroup_free(struct cgroup *cgrp); -void cgroup_move_task(struct task_struct *p, struct css_set *to); - struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf, size_t nbytes, enum psi_res res, struct kernfs_open_file *of); @@ -39,6 +33,11 @@ void psi_trigger_destroy(struct psi_trigger *t); __poll_t psi_trigger_poll(void **trigger_ptr, struct file *file, poll_table *wait); + +#ifdef CONFIG_CGROUPS +int psi_cgroup_alloc(struct cgroup *cgrp); +void psi_cgroup_free(struct cgroup *cgrp); +void cgroup_move_task(struct task_struct *p, struct css_set *to); #endif #else /* CONFIG_PSI */ -- 2.34.1

2 1

[PATCH OLK-5.10 00/38] Add error handling for device_add_disk
by Zhong Jinghua 25 Oct '23

25 Oct '23

Befor these paths, block is not know when drvier is add fail. Now, add error handling for drivers. Christoph Hellwig (14): block: fold register_disk into device_add_disk block: call blk_integrity_add earlier in device_add_disk block: add the events* attributes to disk_attrs block: fix error unwinding in device_add_disk block: clear ->slave_dir when dropping the main slave_dir reference block: add blk_alloc_disk and blk_cleanup_disk APIs blk-mq: add the blk_mq_alloc_disk APIs loop: use blk_mq_alloc_disk and blk_cleanup_disk loop: fix order of cleaning up the queue and freeing the tagset nbd: use blk_mq_alloc_disk and blk_cleanup_disk block: add a flag to make put_disk on partially initalized disks safer nvme: use blk_mq_alloc_disk brd: convert to blk_alloc_disk/blk_cleanup_disk ubi: use blk_mq_alloc_disk and blk_cleanup_disk Dan Carpenter (1): blk-mq: fix an IS_ERR() vs NULL bug Luis Chamberlain (11): block: return errors from blk_integrity_add block: return errors from disk_alloc_events block: add error handling for device_add_disk / add_disk block: fix device_add_disk() kobject_create_and_add() error handling loop: add error handling support for add_disk() nbd: add error handling support for add_disk() nvme: add error handling support for add_disk() block/brd: add error handling support for add_disk() scsi: sd: Add error handling support for add_disk() scsi: sr: Add error handling support for add_disk() mtd/ubi/block: add error handling support for add_disk() Tetsuo Handa (1): block: check minor range in device_add_disk() Wang Qing (1): nbd: fix order of cleaning up the queue and freeing the tagset Yu Kuai (1): block: fix memory leak for elevator on add_disk failure Zhong Jinghua (9): block: return errors from blk_register_region block: Fix the kabi change in device_add_disk block: Fix the kabi change on blk_register_region block: call blk_get_queue earlier in __device_add_disk block: Fix minor range check in device_add_disk() block: Set memalloc_noio to false in the error path scsi: sd: Clean up sdkp if device_add_disk() failed mtd/ubi/block: Fix null pointer dereference issue in error path mtd/ubi/block: Fix uaf problem in ubiblock_cleanup block/blk-integrity.c | 12 +- block/blk-mq.c | 19 +++ block/blk.h | 5 +- block/genhd.c | 286 ++++++++++++++++++++++++++------------- drivers/block/brd.c | 103 ++++++-------- drivers/block/loop.c | 26 ++-- drivers/block/nbd.c | 59 ++++---- drivers/mtd/ubi/block.c | 76 +++++------ drivers/nvme/host/core.c | 41 +++--- drivers/scsi/sd.c | 8 +- drivers/scsi/sr.c | 7 +- include/linux/blk-mq.h | 12 ++ include/linux/genhd.h | 36 +++++ 13 files changed, 417 insertions(+), 273 deletions(-) -- 2.31.1

2 39

[openEuler/rdma-core] libhns: Support flexible WQE buffer page size
by Chengchang Tang 23 Oct '23

23 Oct '23

driver inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/I87LTM -------------------------------------------------------------------------- Currently, driver fixedly allocates 4K pages for user space WQE buffer even in a 64K system. This results in HW reading WQE with a granularity of 4K even in a 64K system. Considering that we support 1024-byte inline, in the scenario of using SQ inline, HW will switch pages every 4 WQEs. This will introduce a delay of about 400ns, which is an average delay of 100ns per packet. In order to improve performance, we allow user-mode drivers to use a larger page size to allocate WQE buffers, thereby reducing the latency introduced by HW page switching. User-mode drivers will be allowed to allocate WQE buffers between 4K to system page size. During ibv_create_qp(), the driver will dynamically select the appropriate page size based on ibv_qp_cap, thus reducing memory consumption while improving performance. This feature needs to be used in conjunction with the kernel-mode driver. In order to ensure forward compatibility, if the kernel-mode driver does not support this feature, the user-mode driver will continue to use a fixed 4K pagesize to allocate WQE buffer. Signed-off-by: Chengchang Tang <tangchengchang(a)huawei.com> --- kernel-headers/rdma/hns-abi.h | 5 ++- providers/hns/hns_roce_u.c | 2 +- providers/hns/hns_roce_u.h | 1 + providers/hns/hns_roce_u_verbs.c | 65 ++++++++++++++++++++++++++------ 4 files changed, 59 insertions(+), 14 deletions(-) diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h index cab941f..157dc9d 100644 --- a/kernel-headers/rdma/hns-abi.h +++ b/kernel-headers/rdma/hns-abi.h @@ -81,7 +81,8 @@ struct hns_roce_ib_create_qp { __u8 log_sq_bb_count; __u8 log_sq_stride; __u8 sq_no_prefetch; - __u8 reserved[5]; + __u8 reserved[4]; + __u8 pageshift; __aligned_u64 sdb_addr; __aligned_u64 comp_mask; __aligned_u64 create_flags; @@ -122,6 +123,7 @@ enum { HNS_ROCE_RQ_INLINE_FLAGS = 1 << 1, HNS_ROCE_CQE_INLINE_FLAGS = 1 << 2, HNS_ROCE_UCTX_CONFIG_DCA = 1 << 3, + HNS_ROCE_UCTX_DYN_QP_PGSZ = 1 << 4, }; enum { @@ -129,6 +131,7 @@ enum { HNS_ROCE_RSP_RQ_INLINE_FLAGS = 1 << 1, HNS_ROCE_RSP_CQE_INLINE_FLAGS = 1 << 2, HNS_ROCE_UCTX_RSP_DCA_FLAGS = HNS_ROCE_UCTX_CONFIG_DCA, + HNS_ROCE_UCTX_RSP_DYN_QP_PGSZ = HNS_ROCE_UCTX_DYN_QP_PGSZ, }; struct hns_roce_ib_alloc_ucontext_resp { diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c index 0660081..02ad880 100644 --- a/providers/hns/hns_roce_u.c +++ b/providers/hns/hns_roce_u.c @@ -267,7 +267,7 @@ static void ucontext_set_cmd(struct hns_roce_alloc_ucontext *cmd, struct hnsdv_context_attr *attr) { cmd->config |= HNS_ROCE_EXSGE_FLAGS | HNS_ROCE_RQ_INLINE_FLAGS | - HNS_ROCE_CQE_INLINE_FLAGS; + HNS_ROCE_CQE_INLINE_FLAGS | HNS_ROCE_UCTX_DYN_QP_PGSZ; if (!attr || !(attr->flags & HNSDV_CONTEXT_FLAGS_DCA)) return; diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h index 5501d8e..ae9ae51 100644 --- a/providers/hns/hns_roce_u.h +++ b/providers/hns/hns_roce_u.h @@ -409,6 +409,7 @@ struct hns_roce_qp { uint8_t sl; uint8_t tc_mode; uint8_t priority; + uint8_t pageshift; unsigned int qkey; enum ibv_mtu path_mtu; diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c index 7b58dd0..f76341c 100644 --- a/providers/hns/hns_roce_u_verbs.c +++ b/providers/hns/hns_roce_u_verbs.c @@ -1327,31 +1327,69 @@ static void free_recv_rinl_buf(struct hns_roce_rinl_buf *rinl_buf) } } +static void get_best_multi_region_pg_shift(struct hns_roce_device *hr_dev, + struct hns_roce_context *ctx, + struct hns_roce_qp *qp, bool dca_en) +{ + uint32_t ext_sge_size; + uint32_t sq_size; + uint32_t rq_size; + uint8_t pg_shift; + + if (!(ctx->config & HNS_ROCE_UCTX_RSP_DYN_QP_PGSZ) || dca_en) { + qp->pageshift = HNS_HW_PAGE_SHIFT; + return; + } + + /* + * The larger the pagesize used, the better the performance, but it + * may waste more memory. Therefore, we use the least common multiple + * (aligned to power of 2) of sq wqe buffer size, rq wqe buffer size, + * and ext_sge buffer size as the pagesize. Additionally, since the + * kernel cannot guarantee the allocation of contiguous memory larger + * than the system page, the pagesize must be smaller than the system + * page. + */ + sq_size = qp->sq.wqe_cnt << qp->sq.wqe_shift; + ext_sge_size = qp->ex_sge.sge_cnt << qp->ex_sge.sge_shift; + rq_size = qp->rq.wqe_cnt << qp->rq.wqe_shift; + + pg_shift = max_t(uint8_t, sq_size ? hr_ilog32(sq_size) : 0, + ext_sge_size ? hr_ilog32(ext_sge_size) : 0); + pg_shift = max_t(uint8_t, pg_shift, rq_size ? hr_ilog32(rq_size) : 0); + pg_shift = max_t(uint8_t, pg_shift, HNS_HW_PAGE_SHIFT); + qp->pageshift = min_t(uint8_t, pg_shift, hr_ilog32(hr_dev->page_size)); +} + static int calc_qp_buff_size(struct hns_roce_device *hr_dev, - struct hns_roce_qp *qp) + struct hns_roce_context *ctx, + struct hns_roce_qp *qp, bool dca_en) { struct hns_roce_wq *sq = &qp->sq; struct hns_roce_wq *rq = &qp->rq; + unsigned int page_size; unsigned int size; qp->buf_size = 0; + get_best_multi_region_pg_shift(hr_dev, ctx, qp, dca_en); + page_size = 1 << qp->pageshift; /* SQ WQE */ sq->offset = 0; - size = to_hr_hem_entries_size(sq->wqe_cnt, sq->wqe_shift); + size = align(sq->wqe_cnt << sq->wqe_shift, page_size); qp->buf_size += size; /* extend SGE WQE in SQ */ qp->ex_sge.offset = qp->buf_size; if (qp->ex_sge.sge_cnt > 0) { - size = to_hr_hem_entries_size(qp->ex_sge.sge_cnt, - qp->ex_sge.sge_shift); + size = align(qp->ex_sge.sge_cnt << qp->ex_sge.sge_shift, + page_size); qp->buf_size += size; } /* RQ WQE */ rq->offset = qp->buf_size; - size = to_hr_hem_entries_size(rq->wqe_cnt, rq->wqe_shift); + size = align(rq->wqe_cnt << rq->wqe_shift, page_size); qp->buf_size += size; if (qp->buf_size < 1) @@ -1375,7 +1413,7 @@ static inline bool check_qp_support_dca(struct hns_roce_dca_ctx *dca_ctx, if (hns_attr && (hns_attr->comp_mask & HNSDV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS) && (hns_attr->create_flags & HNSDV_QP_CREATE_ENABLE_DCA_MODE)) - return true; + return dca_ctx->max_size > 0; return false; } @@ -1396,9 +1434,12 @@ static int qp_alloc_wqe(struct ibv_qp_init_attr_ex *attr, struct hns_roce_qp *qp, struct hns_roce_context *ctx) { struct hns_roce_device *hr_dev = to_hr_dev(ctx->ibv_ctx.context.device); + bool dca_en = check_qp_support_dca(&ctx->dca_ctx, attr, hns_attr); + int ret; - if (calc_qp_buff_size(hr_dev, qp)) - return -EINVAL; + ret = calc_qp_buff_size(hr_dev, ctx, qp, dca_en); + if (ret) + return ret; qp->sq.wrid = malloc(qp->sq.wqe_cnt * sizeof(uint64_t)); if (!qp->sq.wrid) @@ -1416,19 +1457,18 @@ static int qp_alloc_wqe(struct ibv_qp_init_attr_ex *attr, goto err_alloc; } - if (check_qp_support_dca(&ctx->dca_ctx, attr, hns_attr) && - ctx->dca_ctx.max_size > 0) { + if (dca_en) { /* when DCA is enabled, use a buffer list to store page addr */ qp->buf.buf = NULL; qp->dca_wqe.max_cnt = hr_hw_page_count(qp->buf_size); - qp->dca_wqe.shift = HNS_HW_PAGE_SHIFT; + qp->dca_wqe.shift = qp->pageshift; qp->dca_wqe.bufs = calloc(qp->dca_wqe.max_cnt, sizeof(void *)); if (!qp->dca_wqe.bufs) goto err_alloc; verbs_debug(&ctx->ibv_ctx, "alloc DCA buf.\n"); } else { if (hns_roce_alloc_buf(&qp->buf, qp->buf_size, - HNS_HW_PAGE_SIZE)) + 1 << qp->pageshift)) goto err_alloc; } @@ -1642,6 +1682,7 @@ static int qp_exec_create_cmd(struct ibv_qp_init_attr_ex *attr, cmd_ex.buf_addr = (uintptr_t)qp->buf.buf; cmd_ex.log_sq_stride = qp->sq.wqe_shift; cmd_ex.log_sq_bb_count = hr_ilog32(qp->sq.wqe_cnt); + cmd_ex.pageshift = qp->pageshift; if (cmd_flag->congest_type_flags) { cmd_ex.comp_mask |= HNS_ROCE_CREATE_QP_MASK_CONGEST_TYPE; -- 2.30.0

1 2