[PATCH v2] accel/habanalabs: Make use of rhashtable
Cai Huoqing
cai.huoqing at linux.dev
Mon May 8 05:28:43 UTC 2023
On 07 5月 23 16:17:55, Oded Gabbay wrote:
> On Sat, May 6, 2023 at 12:25 PM Cai Huoqing <cai.huoqing at linux.dev> wrote:
> >
> > On 04 5月 23 09:12:40, Oded Gabbay wrote:
> > > On Thu, May 4, 2023 at 6:00 AM Cai Huoqing <cai.huoqing at linux.dev> wrote:
> > > >
> > > > On 30 4月 23 09:36:29, Oded Gabbay wrote:
> > > > > On Fri, Apr 28, 2023 at 5:49 PM Cai Huoqing <cai.huoqing at linux.dev> wrote:
> > > > > >
> > > > > > Using rhashtable to accelerate the search for userptr by address,
> > > > > > instead of using a list.
> > > > > >
> > > > > > Preferably, the lookup complexity of a hash table is O(1).
> > > > > >
> > > > > > This patch will speedup the method
> > > > > > hl_userptr_is_pinned by rhashtable_lookup_fast.
> > > > > >
> > > > > > Signed-off-by: Cai Huoqing <cai.huoqing at linux.dev>
> > > > >
> > > > > Thanks for the patch, but the reason we never optimized this path is
> > > > > because this code path is only relevant for Goya, which we don't want
> > > > > to change the code for anymore.
> > > > > For Gaudi we don't pin the memory in the host during submission. It is
> > > > > done much earlier, when the user maps the memory to the device. The
> > > > > code path in Gaudi is only in case the device is loaded with its MMU
> > > > > disabled. This mode was used only for debug/bring-up of the ASIC many
> > > > > years ago. As you can see in Gaudi2, that mode was dropped even for
> > > >
> > > > Do you mean that the userspace directly call HL_MEM_OP_MAP/HL_MEM_OP_UNMAP
> > > > with a flag HL_MEM_USERPTR instead of pin host mem in submissmion?
> > > Yes, correct.
> >
> > BTW,
> >
> > Another question,
> >
> > I found the PMMU in the driver code, does it support for like CudaHostAlloc
> > and hostMap?
> In Habana, you allocate host memory by yourself, not through the
> driver. Just do malloc in your application.
> Then you need to map it to the device to get a virtual address that
> the device is familiar with.
> This is done by calling the MEMORY ioctl with the MAP opcode.
>
> >
> > gaudi seems can access host memory by pcie, some kernel can compute data
> > in host directly, right?
> In theory, yes. However, we don't do that. We move data into the HBM
> and then the kernels work on the data.
> Oded
Agree, but What is the purpose of designing pmmu,
do you think using the pagetable in PMMU can speed up memcpy
H2D/D2H or some memory notity between CPU and device ?
Thanks
Cai-
> >
> > But accessing via PCIE(64GB/s), compared to accessing HBM(3T/s),
> > has too low bandwidth.
> >
> > for nvidia, it has NVLINK.
> >
> > Thank,
> > Cai-
> >
> > >
> > > >
> > > > > debug/bring-up.
> > > > >
> > > > > Therefore, I prefer not to take this patch as validation for both
> > > > > functionality and performance will take time which will be better
> > > > > spent elsewhere.
> > > > >
> > > > > Thanks,
> > > > > Oded
> > > > >
> > > > > > ---
> > > > > > v1->v2:
> > > > > > Use rhashtable_free_and_destroy in hl_userptr_delete_list.
> > > > > >
> > > > > > .../habanalabs/common/command_submission.c | 16 ++++++--
> > > > > > drivers/accel/habanalabs/common/habanalabs.h | 19 +++++----
> > > > > > drivers/accel/habanalabs/common/memory.c | 39 +++++++++++--------
> > > > > > drivers/accel/habanalabs/gaudi/gaudi.c | 16 +++++---
> > > > > > drivers/accel/habanalabs/goya/goya.c | 14 ++++---
> > > > > > 5 files changed, 65 insertions(+), 39 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/accel/habanalabs/common/command_submission.c b/drivers/accel/habanalabs/common/command_submission.c
> > > > > > index af9d2e22c6e7..35c2ab934396 100644
> > > > > > --- a/drivers/accel/habanalabs/common/command_submission.c
> > > > > > +++ b/drivers/accel/habanalabs/common/command_submission.c
> > > > > > @@ -312,7 +312,7 @@ static int cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
> > > > > > parser.job_id = job->id;
> > > > > >
> > > > > > parser.hw_queue_id = job->hw_queue_id;
> > > > > > - parser.job_userptr_list = &job->userptr_list;
> > > > > > + parser.job_userptr_ht = &job->userptr_ht;
> > > > > > parser.patched_cb = NULL;
> > > > > > parser.user_cb = job->user_cb;
> > > > > > parser.user_cb_size = job->user_cb_size;
> > > > > > @@ -351,7 +351,7 @@ static void hl_complete_job(struct hl_device *hdev, struct hl_cs_job *job)
> > > > > > struct hl_cs *cs = job->cs;
> > > > > >
> > > > > > if (is_cb_patched(hdev, job)) {
> > > > > > - hl_userptr_delete_list(hdev, &job->userptr_list);
> > > > > > + hl_userptr_delete_list(hdev, &job->userptr_ht);
> > > > > >
> > > > > > /*
> > > > > > * We might arrive here from rollback and patched CB wasn't
> > > > > > @@ -1284,6 +1284,7 @@ struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev,
> > > > > > enum hl_queue_type queue_type, bool is_kernel_allocated_cb)
> > > > > > {
> > > > > > struct hl_cs_job *job;
> > > > > > + int rc;
> > > > > >
> > > > > > job = kzalloc(sizeof(*job), GFP_ATOMIC);
> > > > > > if (!job)
> > > > > > @@ -1296,13 +1297,20 @@ struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev,
> > > > > > job->queue_type = queue_type;
> > > > > > job->is_kernel_allocated_cb = is_kernel_allocated_cb;
> > > > > >
> > > > > > - if (is_cb_patched(hdev, job))
> > > > > > - INIT_LIST_HEAD(&job->userptr_list);
> > > > > > + if (is_cb_patched(hdev, job)) {
> > > > > > + rc = rhashtable_init(&job->userptr_ht, &hl_userptr_rht_params);
> > > > > > + if (rc)
> > > > > > + goto free_job;
> > > > > > + }
> > > > > >
> > > > > > if (job->queue_type == QUEUE_TYPE_EXT)
> > > > > > INIT_WORK(&job->finish_work, job_wq_completion);
> > > > > >
> > > > > > return job;
> > > > > > +
> > > > > > +free_job:
> > > > > > + kfree(job);
> > > > > > + return NULL;
> > > > > > }
> > > > > >
> > > > > > static enum hl_cs_type hl_cs_get_cs_type(u32 cs_type_flags)
> > > > > > diff --git a/drivers/accel/habanalabs/common/habanalabs.h b/drivers/accel/habanalabs/common/habanalabs.h
> > > > > > index eaae69a9f817..9c876d1480d2 100644
> > > > > > --- a/drivers/accel/habanalabs/common/habanalabs.h
> > > > > > +++ b/drivers/accel/habanalabs/common/habanalabs.h
> > > > > > @@ -19,6 +19,7 @@
> > > > > > #include <linux/dma-direction.h>
> > > > > > #include <linux/scatterlist.h>
> > > > > > #include <linux/hashtable.h>
> > > > > > +#include <linux/rhashtable.h>
> > > > > > #include <linux/debugfs.h>
> > > > > > #include <linux/rwsem.h>
> > > > > > #include <linux/eventfd.h>
> > > > > > @@ -540,6 +541,8 @@ struct hl_hints_range {
> > > > > > u64 end_addr;
> > > > > > };
> > > > > >
> > > > > > +extern const struct rhashtable_params hl_userptr_rht_params;
> > > > > > +
> > > > > > /**
> > > > > > * struct asic_fixed_properties - ASIC specific immutable properties.
> > > > > > * @hw_queues_props: H/W queues properties.
> > > > > > @@ -1915,7 +1918,7 @@ struct hl_ctx_mgr {
> > > > > > /**
> > > > > > * struct hl_userptr - memory mapping chunk information
> > > > > > * @vm_type: type of the VM.
> > > > > > - * @job_node: linked-list node for hanging the object on the Job's list.
> > > > > > + * @job_node: hashtable node for hanging the object on the Job's list.
> > > > > > * @pages: pointer to struct page array
> > > > > > * @npages: size of @pages array
> > > > > > * @sgt: pointer to the scatter-gather table that holds the pages.
> > > > > > @@ -1928,7 +1931,7 @@ struct hl_ctx_mgr {
> > > > > > */
> > > > > > struct hl_userptr {
> > > > > > enum vm_type vm_type; /* must be first */
> > > > > > - struct list_head job_node;
> > > > > > + struct rhash_head job_node;
> > > > > > struct page **pages;
> > > > > > unsigned int npages;
> > > > > > struct sg_table *sgt;
> > > > > > @@ -2028,7 +2031,7 @@ struct hl_cs {
> > > > > > * @patched_cb: in case of patching, this is internal CB which is submitted on
> > > > > > * the queue instead of the CB we got from the IOCTL.
> > > > > > * @finish_work: workqueue object to run when job is completed.
> > > > > > - * @userptr_list: linked-list of userptr mappings that belong to this job and
> > > > > > + * @userptr_ht: hashtable of userptr mappings that belong to this job and
> > > > > > * wait for completion.
> > > > > > * @debugfs_list: node in debugfs list of command submission jobs.
> > > > > > * @refcount: reference counter for usage of the CS job.
> > > > > > @@ -2056,7 +2059,7 @@ struct hl_cs_job {
> > > > > > struct hl_cb *user_cb;
> > > > > > struct hl_cb *patched_cb;
> > > > > > struct work_struct finish_work;
> > > > > > - struct list_head userptr_list;
> > > > > > + struct rhashtable userptr_ht;
> > > > > > struct list_head debugfs_list;
> > > > > > struct kref refcount;
> > > > > > enum hl_queue_type queue_type;
> > > > > > @@ -2075,7 +2078,7 @@ struct hl_cs_job {
> > > > > > * @user_cb: the CB we got from the user.
> > > > > > * @patched_cb: in case of patching, this is internal CB which is submitted on
> > > > > > * the queue instead of the CB we got from the IOCTL.
> > > > > > - * @job_userptr_list: linked-list of userptr mappings that belong to the related
> > > > > > + * @job_userptr_ht: hashtable of userptr mappings that belong to the related
> > > > > > * job and wait for completion.
> > > > > > * @cs_sequence: the sequence number of the related CS.
> > > > > > * @queue_type: the type of the H/W queue this job is submitted to.
> > > > > > @@ -2098,7 +2101,7 @@ struct hl_cs_job {
> > > > > > struct hl_cs_parser {
> > > > > > struct hl_cb *user_cb;
> > > > > > struct hl_cb *patched_cb;
> > > > > > - struct list_head *job_userptr_list;
> > > > > > + struct rhashtable *job_userptr_ht;
> > > > > > u64 cs_sequence;
> > > > > > enum hl_queue_type queue_type;
> > > > > > u32 ctx_id;
> > > > > > @@ -3760,9 +3763,9 @@ int hl_pin_host_memory(struct hl_device *hdev, u64 addr, u64 size,
> > > > > > struct hl_userptr *userptr);
> > > > > > void hl_unpin_host_memory(struct hl_device *hdev, struct hl_userptr *userptr);
> > > > > > void hl_userptr_delete_list(struct hl_device *hdev,
> > > > > > - struct list_head *userptr_list);
> > > > > > + struct rhashtable *userptr_ht);
> > > > > > bool hl_userptr_is_pinned(struct hl_device *hdev, u64 addr, u32 size,
> > > > > > - struct list_head *userptr_list,
> > > > > > + struct rhashtable *userptr_ht,
> > > > > > struct hl_userptr **userptr);
> > > > > >
> > > > > > int hl_mmu_init(struct hl_device *hdev);
> > > > > > diff --git a/drivers/accel/habanalabs/common/memory.c b/drivers/accel/habanalabs/common/memory.c
> > > > > > index a7b6a273ce21..fa2104e33639 100644
> > > > > > --- a/drivers/accel/habanalabs/common/memory.c
> > > > > > +++ b/drivers/accel/habanalabs/common/memory.c
> > > > > > @@ -23,6 +23,13 @@ MODULE_IMPORT_NS(DMA_BUF);
> > > > > >
> > > > > > #define MEM_HANDLE_INVALID ULONG_MAX
> > > > > >
> > > > > > +const struct rhashtable_params hl_userptr_rht_params = {
> > > > > > + .head_offset = offsetof(struct hl_userptr, job_node),
> > > > > > + .key_offset = offsetof(struct hl_userptr, addr),
> > > > > > + .key_len = sizeof(u64),
> > > > > > + .automatic_shrinking = true,
> > > > > > +};
> > > > > > +
> > > > > > static int allocate_timestamps_buffers(struct hl_fpriv *hpriv,
> > > > > > struct hl_mem_in *args, u64 *handle);
> > > > > >
> > > > > > @@ -2483,7 +2490,6 @@ int hl_pin_host_memory(struct hl_device *hdev, u64 addr, u64 size,
> > > > > > userptr->size = size;
> > > > > > userptr->addr = addr;
> > > > > > userptr->dma_mapped = false;
> > > > > > - INIT_LIST_HEAD(&userptr->job_node);
> > > > > >
> > > > > > rc = get_user_memory(hdev, addr, size, npages, start, offset,
> > > > > > userptr);
> > > > > > @@ -2522,32 +2528,32 @@ void hl_unpin_host_memory(struct hl_device *hdev, struct hl_userptr *userptr)
> > > > > > unpin_user_pages_dirty_lock(userptr->pages, userptr->npages, true);
> > > > > > kvfree(userptr->pages);
> > > > > >
> > > > > > - list_del(&userptr->job_node);
> > > > > > -
> > > > > > sg_free_table(userptr->sgt);
> > > > > > kfree(userptr->sgt);
> > > > > > }
> > > > > >
> > > > > > +static void hl_userptr_free_cb(void *ptr, void *arg)
> > > > > > +{
> > > > > > + struct hl_userptr *userptr = ptr;
> > > > > > + struct hl_device *hdev = (struct hl_device *)arg;
> > > > > > +
> > > > > > + hl_unpin_host_memory(hdev, userptr);
> > > > > > + kfree(userptr);
> > > > > > +}
> > > > > > +
> > > > > > /**
> > > > > > * hl_userptr_delete_list() - clear userptr list.
> > > > > > * @hdev: pointer to the habanalabs device structure.
> > > > > > - * @userptr_list: pointer to the list to clear.
> > > > > > + * @userptr_ht: pointer to the hashtable to clear.
> > > > > > *
> > > > > > * This function does the following:
> > > > > > * - Iterates over the list and unpins the host memory and frees the userptr
> > > > > > * structure.
> > > > > > */
> > > > > > void hl_userptr_delete_list(struct hl_device *hdev,
> > > > > > - struct list_head *userptr_list)
> > > > > > + struct rhashtable *userptr_ht)
> > > > > > {
> > > > > > - struct hl_userptr *userptr, *tmp;
> > > > > > -
> > > > > > - list_for_each_entry_safe(userptr, tmp, userptr_list, job_node) {
> > > > > > - hl_unpin_host_memory(hdev, userptr);
> > > > > > - kfree(userptr);
> > > > > > - }
> > > > > > -
> > > > > > - INIT_LIST_HEAD(userptr_list);
> > > > > > + rhashtable_free_and_destroy(userptr_ht, hl_userptr_free_cb, hdev);
> > > > > > }
> > > > > >
> > > > > > /**
> > > > > > @@ -2555,7 +2561,7 @@ void hl_userptr_delete_list(struct hl_device *hdev,
> > > > > > * @hdev: pointer to the habanalabs device structure.
> > > > > > * @addr: user address to check.
> > > > > > * @size: user block size to check.
> > > > > > - * @userptr_list: pointer to the list to clear.
> > > > > > + * @userptr_ht: pointer to the hashtable to clear.
> > > > > > * @userptr: pointer to userptr to check.
> > > > > > *
> > > > > > * This function does the following:
> > > > > > @@ -2563,10 +2569,11 @@ void hl_userptr_delete_list(struct hl_device *hdev,
> > > > > > * pinned. If so, returns true, otherwise returns false.
> > > > > > */
> > > > > > bool hl_userptr_is_pinned(struct hl_device *hdev, u64 addr,
> > > > > > - u32 size, struct list_head *userptr_list,
> > > > > > + u32 size, struct rhashtable *userptr_ht,
> > > > > > struct hl_userptr **userptr)
> > > > > > {
> > > > > > - list_for_each_entry((*userptr), userptr_list, job_node) {
> > > > > > + (*userptr) = rhashtable_lookup_fast(userptr_ht, &addr, hl_userptr_rht_params);
> > > > > > + if (*userptr) {
> > > > > > if ((addr == (*userptr)->addr) && (size == (*userptr)->size))
> > > > > > return true;
> > > > > > }
> > > > > > diff --git a/drivers/accel/habanalabs/gaudi/gaudi.c b/drivers/accel/habanalabs/gaudi/gaudi.c
> > > > > > index a29aa8f7b6f3..1e1433042413 100644
> > > > > > --- a/drivers/accel/habanalabs/gaudi/gaudi.c
> > > > > > +++ b/drivers/accel/habanalabs/gaudi/gaudi.c
> > > > > > @@ -1031,7 +1031,7 @@ static int _gaudi_init_tpc_mem(struct hl_device *hdev,
> > > > > > }
> > > > > >
> > > > > > free_job:
> > > > > > - hl_userptr_delete_list(hdev, &job->userptr_list);
> > > > > > + hl_userptr_delete_list(hdev, &job->userptr_ht);
> > > > > > hl_debugfs_remove_job(hdev, job);
> > > > > > kfree(job);
> > > > > > atomic_dec(&cb->cs_cnt);
> > > > > > @@ -4901,7 +4901,7 @@ static int gaudi_pin_memory_before_cs(struct hl_device *hdev,
> > > > > > int rc;
> > > > > >
> > > > > > if (hl_userptr_is_pinned(hdev, addr, le32_to_cpu(user_dma_pkt->tsize),
> > > > > > - parser->job_userptr_list, &userptr))
> > > > > > + parser->job_userptr_ht, &userptr))
> > > > > > goto already_pinned;
> > > > > >
> > > > > > userptr = kzalloc(sizeof(*userptr), GFP_KERNEL);
> > > > > > @@ -4913,7 +4913,10 @@ static int gaudi_pin_memory_before_cs(struct hl_device *hdev,
> > > > > > if (rc)
> > > > > > goto free_userptr;
> > > > > >
> > > > > > - list_add_tail(&userptr->job_node, parser->job_userptr_list);
> > > > > > + rc = rhashtable_insert_fast(parser->job_userptr_ht,
> > > > > > + &userptr->job_node, hl_userptr_rht_params);
> > > > > > + if (rc)
> > > > > > + goto unpin_memory;
> > > > > >
> > > > > > rc = hdev->asic_funcs->asic_dma_map_sgtable(hdev, userptr->sgt, dir);
> > > > > > if (rc) {
> > > > > > @@ -4931,7 +4934,8 @@ static int gaudi_pin_memory_before_cs(struct hl_device *hdev,
> > > > > > return 0;
> > > > > >
> > > > > > unpin_memory:
> > > > > > - list_del(&userptr->job_node);
> > > > > > + rhashtable_remove_fast(parser->job_userptr_ht,
> > > > > > + &userptr->job_node, hl_userptr_rht_params);
> > > > > > hl_unpin_host_memory(hdev, userptr);
> > > > > > free_userptr:
> > > > > > kfree(userptr);
> > > > > > @@ -5175,7 +5179,7 @@ static int gaudi_patch_dma_packet(struct hl_device *hdev,
> > > > > > if ((!skip_host_mem_pin) &&
> > > > > > (!hl_userptr_is_pinned(hdev, addr,
> > > > > > le32_to_cpu(user_dma_pkt->tsize),
> > > > > > - parser->job_userptr_list, &userptr))) {
> > > > > > + parser->job_userptr_ht, &userptr))) {
> > > > > > dev_err(hdev->dev, "Userptr 0x%llx + 0x%x NOT mapped\n",
> > > > > > addr, user_dma_pkt->tsize);
> > > > > > return -EFAULT;
> > > > > > @@ -5472,7 +5476,7 @@ static int gaudi_parse_cb_no_mmu(struct hl_device *hdev,
> > > > > >
> > > > > > free_userptr:
> > > > > > if (rc)
> > > > > > - hl_userptr_delete_list(hdev, parser->job_userptr_list);
> > > > > > + hl_userptr_delete_list(hdev, parser->job_userptr_ht);
> > > > > > return rc;
> > > > > > }
> > > > > >
> > > > > > diff --git a/drivers/accel/habanalabs/goya/goya.c b/drivers/accel/habanalabs/goya/goya.c
> > > > > > index fb0ac9df841a..bfcbb9e8b126 100644
> > > > > > --- a/drivers/accel/habanalabs/goya/goya.c
> > > > > > +++ b/drivers/accel/habanalabs/goya/goya.c
> > > > > > @@ -3347,7 +3347,7 @@ static int goya_pin_memory_before_cs(struct hl_device *hdev,
> > > > > > int rc;
> > > > > >
> > > > > > if (hl_userptr_is_pinned(hdev, addr, le32_to_cpu(user_dma_pkt->tsize),
> > > > > > - parser->job_userptr_list, &userptr))
> > > > > > + parser->job_userptr_ht, &userptr))
> > > > > > goto already_pinned;
> > > > > >
> > > > > > userptr = kzalloc(sizeof(*userptr), GFP_KERNEL);
> > > > > > @@ -3359,7 +3359,10 @@ static int goya_pin_memory_before_cs(struct hl_device *hdev,
> > > > > > if (rc)
> > > > > > goto free_userptr;
> > > > > >
> > > > > > - list_add_tail(&userptr->job_node, parser->job_userptr_list);
> > > > > > + rc = rhashtable_insert_fast(parser->job_userptr_ht,
> > > > > > + &userptr->job_node, hl_userptr_rht_params);
> > > > > > + if (rc)
> > > > > > + goto unpin_memory;
> > > > > >
> > > > > > rc = hdev->asic_funcs->asic_dma_map_sgtable(hdev, userptr->sgt, dir);
> > > > > > if (rc) {
> > > > > > @@ -3377,7 +3380,8 @@ static int goya_pin_memory_before_cs(struct hl_device *hdev,
> > > > > > return 0;
> > > > > >
> > > > > > unpin_memory:
> > > > > > - list_del(&userptr->job_node);
> > > > > > + rhashtable_remove_fast(parser->job_userptr_ht,
> > > > > > + &userptr->job_node, hl_userptr_rht_params);
> > > > > > hl_unpin_host_memory(hdev, userptr);
> > > > > > free_userptr:
> > > > > > kfree(userptr);
> > > > > > @@ -3806,7 +3810,7 @@ static int goya_patch_dma_packet(struct hl_device *hdev,
> > > > > > if ((!skip_host_mem_pin) &&
> > > > > > (hl_userptr_is_pinned(hdev, addr,
> > > > > > le32_to_cpu(user_dma_pkt->tsize),
> > > > > > - parser->job_userptr_list, &userptr) == false)) {
> > > > > > + parser->job_userptr_ht, &userptr) == false)) {
> > > > > > dev_err(hdev->dev, "Userptr 0x%llx + 0x%x NOT mapped\n",
> > > > > > addr, user_dma_pkt->tsize);
> > > > > > return -EFAULT;
> > > > > > @@ -4104,7 +4108,7 @@ static int goya_parse_cb_no_mmu(struct hl_device *hdev,
> > > > > >
> > > > > > free_userptr:
> > > > > > if (rc)
> > > > > > - hl_userptr_delete_list(hdev, parser->job_userptr_list);
> > > > > > + hl_userptr_delete_list(hdev, parser->job_userptr_ht);
> > > > > > return rc;
> > > > > > }
> > > > > >
> > > > > > --
> > > > > > 2.34.1
> > > > > >
More information about the dri-devel
mailing list