On 2022-03-01 23:29, Miles Chen via iommu wrote:

Hi Yunfei,

...

...

Since __alloc_and_insert_iova_range fail will set the current alloc iova size to max32_alloc_size (iovad->max32_alloc_size = size), when the retry is executed into the __alloc_and_insert_iova_range function, the retry action will be blocked by the check condition (size >= iovad->max32_alloc_size) and goto iova32_full directly, causes the action of retry regular alloc iova in __alloc_and_insert_iova_range to not actually be executed.

Based on the above, so need reset max32_alloc_size before retry alloc iova when alloc iova fail, that is set the initial dma_32bit_pfn value of iovad to max32_alloc_size, so that the action of retry alloc iova in __alloc_and_insert_iova_range can be executed.

Have you observed this making any difference in practice?

Given that both free_cpu_cached_iovas() and free_global_cached_iovas() call iova_magazine_free_pfns(), which calls remove_iova(), which calls __cached_rbnode_delete_update(), I'm thinking no...

Robin.

Like Robin pointed out, if some cached iovas are freed by free_global_cached_iovas()/free_cpu_cached_iovas(), the max32_alloc_size should be reset to iovad->dma_32bit_pfn.

If no cached iova is freed, resetting max32_alloc_size before the retry allocation only give us a retry. Is it possible that other users free their iovas during the additional retry?

No, it's not possible, since everyone's serialised by iova_rbtree_lock. If the caches were already empty and the retry gets the lock first, it will still fail again - forcing a reset of max32_alloc_size only means it has to take the slow path to that failure. If another caller *did* manage to get in and free something between free_global_cached_iovas() dropping the lock and alloc_iova() re-taking it, then that would have legitimately reset max32_alloc_size anyway.

Thanks, Robin.

alloc_iova_fast() retry: alloc_iova() // failed, iovad->max32_alloc_size = size free_cpu_cached_iovas() iova_magazine_free_pfns() remove_iova() __cached_rbnode_delete_update() iovad->max32_alloc_size = iovad->dma_32bit_pfn // reset free_global_cached_iovas() iova_magazine_free_pfns() remove_iova() __cached_rbnode_delete_update() iovad->max32_alloc_size = iovad->dma_32bit_pfn // reset goto retry;

thanks, Miles _______________________________________________ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu

From: yf.wang@mediatek.com yf.wang@medaitek.com

On Thu, 2022-03-03 at 07:52 +0800, Miles Chen wrote:

Thanks for your explanation.

YF showed me some numbers yesterday and maybe we can have a further discussion in that test case. (It looks like that some iovas are freed but their pfn_lo(s) are less than cached_iova->pfn_lo, so max32_alloc_size is not reset. So there are enought free iovas but the allocation still failed)

__cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free) { struct iova *cached_iova;

  cached_iova = to_iova(iovad->cached32_node);
  if (free == cached_iova ||
          (free->pfn_hi < iovad->dma_32bit_pfn &&
           free->pfn_lo >= cached_iova->pfn_lo)) {
          iovad->cached32_node = rb_next(&free->node);
          iovad->max32_alloc_size = iovad->dma_32bit_pfn;
  }
  ...

}

Hi YF, Could your share your observation of the iova allocation failure you hit?

Thanks, Miles

Hi Miles & Robin,

The scenario we encountered in our actual test: An iova domain with a 3GB iova region (0x4000_0000~0x1_0000_0000), only alloc iova total size 614788KB), then alloc iova size=0x731f000 fail, and then the retry mechanism is free 300 cache iova at first, and then retry still fail, the reason for fail is blocked by the check condition (size >= iovad->max32_alloc_size) and goto iova32_full directly.

Why is no reset max32_alloc_size in __cached_rbnode_delete_update? Because of the pfn_lo of the freed cache iova is smaller than the current cached32_node's pfn_lo (free->pfn_lo < cached_iova->pfn_lo), so will not execute (iovad->max32_alloc_size = iovad->dma_32bit_pfn), is blocked by the check condition (free->pfn_lo >= cached_iova-

pfn_lo).

The issue observed during stress testing, when alloc fail reset max32_alloc_size to retry can help to allocate iova to those lower addresses.

The following is the actual stress test log dump: ('pfn_lo', 'size' and 'pfn_hi' has iova_shift in dump log) size: 0x731f000 max32_alloc_size: 0x731f000 cached_iova: 0xffffff814751a1c0 cached_iova->pfn_lo: 0xf3500000 cached_iova->pfn_hi: 0xf35ff000

1.__alloc_and_insert_iova_range_fail [name:iova&]__alloc_and_insert_iova_range_fail: iovad:0xffffff80d5be2808 {granule:0x1000, start_pfn:0x40000, dma_32bit_pfn:0x100000, max32_alloc_size:0x731f},size:0x731f

2.dump all iova nodes of rbtree, contain cached iova: [name:iova&]dump_iova_rbtree start, iovad:0xffffff80d5be2808 {rbroot:0xffffff814751a4c0, cached_node:0xffffff80d5be2860, cached32_node:0xffffff814751a1c0, granule:0x1000, start_pfn:0x40000, dma_32bit_pfn:0x100000, max32_alloc_size:0x731f}

('pfn_lo', 'size' and 'pfn_hi' has iova_shift in dump log) [name:iova&]index iova_object pfn_lo size pfn_hi [name:iova&] 1 0xffffff814751ae00 0x40800000 0x4c9000 0x40cc8000 [name:iova&] 2 0xffffff807c960880 0x40d00000 0xc4000 0x40dc3000 ...322 lines of log are omitted here... [name:iova&]325 0xffffff814751a1c0 0xf3500000 0x100000 0xf35ff000 [name:iova&]326 0xffffff8004552080 0xf3600000 0x100000 0xf36ff000 [name:iova&]327 0xffffff80045524c0 0xf3700000 0xbc000 0xf37bb000 [name:iova&]328 0xffffff814751a700 0xf3800000 0x265000 0xf3a64000 [name:iova&]329 0xffffff8004552180 0xf3c00000 0xa13000 0xf4612000 [name:iova&]330 0xffffff80c1b3eb40 0xf4800000 0xb400000 0xffbff000 [name:iova&]331 0xffffff80d6e0b580 0xffcff000 0x1000 0xffcff000 [name:iova&]332 0xffffff80c8395140 0xffd00000 0x80000 0xffd7f000 [name:iova&]333 0xffffff80c8395000 0xffde0000 0x20000 0xffdff000 [name:iova&]334 0xffffff80c8395380 0xffe00000 0x80000 0xffe7f000 [name:iova&]335 0xffffff80c8395880 0xffec0000 0x20000 0xffedf000 [name:iova&]336 0xffffff80c83957c0 0xffef9000 0x1000 0xffef9000 [name:iova&]337 0xffffff80c83956c0 0xffefa000 0x1000 0xffefa000 [name:iova&]338 0xffffff80c8395f40 0xffefb000 0x1000 0xffefb000 [name:iova&]339 0xffffff80c8395a80 0xffefc000 0x4000 0xffeff000 [name:iova&]340 0xffffff80c1b3e840 0xfff00000 0x100000 0xfffff000 [name:iova&]341 0xffffff80d5be2860 0xfffffffffffff000 0x1000 0xfffffffffffff000 [name:iova&]dump_iova_rbtree done, iovad:0xffffff80d5be2808, count:341, total_size:625876KB

3.alloc fail, flushing rcache and retry. [name:iova&]alloc_iova_fast, flushing rcache and retry, iovad:0xffffff80d5be2808, size:0x731f, limit_pfn:0xfffff

4.retry is executed into the __alloc_and_insert_iova_range function: [name:iova&]__alloc_and_insert_iova_range fail goto iova32_full, iovad:0xffffff80d5be2808 {granule:0x1000, start_pfn:0x40000, dma_32bit_pfn:0x100000, max32_alloc_size:0x731f},size:0x731f

5.dump all iova nodes of rbtree, current caches is already empty: [name:iova&]dump_iova_rbtree start, iovad:0xffffff80d5be2808 {rbroot:0xffffff80fe76da80, cached_node:0xffffff80d5be2860, cached32_node:0xffffff814751a1c0, granule:0x1000, start_pfn:0x40000, dma_32bit_pfn:0x100000, max32_alloc_size:0x731f}

('pfn_lo', 'size' and 'pfn_hi' has iova_shift in dump log) [name:iova&]index iova_object pfn_lo size pfn_hi [name:iova&] 1 0xffffff814751ae00 0x40800000 0x4c9000 0x40cc8000 [name:iova&] 2 0xffffff807c960880 0x40d00000 0xc4000 0x40dc3000 [name:iova&] 3 0xffffff81487ce840 0x40e00000 0x37a000 0x41179000 [name:iova&] 4 0xffffff80fe76d600 0x41200000 0x2400000 0x435ff000 [name:iova&] 5 0xffffff8004552000 0x46800000 0x47e2000 0x4afe1000 [name:iova&] 6 0xffffff807c960a40 0x50200000 0x400000 0x505ff000 [name:iova&] 7 0xffffff818cdb5780 0x50600000 0x265000 0x50864000 [name:iova&] 8 0xffffff814751a440 0x50a00000 0x2ae000 0x50cad000 [name:iova&] 9 0xffffff818cdb5d40 0x50e00000 0x4c9000 0x512c8000 [name:iova&] 10 0xffffff81487cef40 0x51400000 0x1a25000 0x52e24000 [name:iova&] 11 0xffffff818cdb5140 0x53c00000 0x400000 0x53fff000 [name:iova&] 12 0xffffff814751a7c0 0x54000000 0x400000 0x543ff000 [name:iova&] 13 0xffffff80fe76d2c0 0x54500000 0xbc000 0x545bb000 [name:iova&] 14 0xffffff81487ce0c0 0x54600000 0x400000 0x549ff000 [name:iova&] 15 0xffffff80fe76d740 0x54a00000 0xbc000 0x54abb000 [name:iova&] 16 0xffffff8004552440 0x54b00000 0xbc000 0x54bbb000 [name:iova&] 17 0xffffff80fe76db80 0x54c00000 0x47e2000 0x593e1000 [name:iova&] 18 0xffffff818cdb5440 0x65000000 0x400000 0x653ff000 [name:iova&] 19 0xffffff8004552880 0x65400000 0x47e2000 0x69be1000 [name:iova&] 20 0xffffff814751acc0 0x6c7e0000 0x20000 0x6c7ff000 [name:iova&] 21 0xffffff80fe76d700 0x6c800000 0x47e2000 0x70fe1000 //0x70fe1000 - 0xebc00000 = 0x7ac1f000:This free range is large enough [name:iova&] 22 0xffffff80fe76da80 0xebc00000 0x400000 0xebfff000 [name:iova&] 23 0xffffff80045523c0 0xec000000 0x400000 0xec3ff000 [name:iova&] 24 0xffffff8004552780 0xec400000 0x400000 0xec7ff000 [name:iova&] 25 0xffffff814751a1c0 0xf3500000 0x100000 0xf35ff000 [name:iova&] 26 0xffffff8004552080 0xf3600000 0x100000 0xf36ff000 [name:iova&] 27 0xffffff80045524c0 0xf3700000 0xbc000 0xf37bb000 [name:iova&] 28 0xffffff814751a700 0xf3800000 0x265000 0xf3a64000 [name:iova&] 29 0xffffff8004552180 0xf3c00000 0xa13000 0xf4612000 [name:iova&] 30 0xffffff80c1b3eb40 0xf4800000 0xb400000 0xffbff000 [name:iova&] 31 0xffffff80d6e0b580 0xffcff000 0x1000 0xffcff000 [name:iova&] 32 0xffffff80c8395140 0xffd00000 0x80000 0xffd7f000 [name:iova&] 33 0xffffff80c8395000 0xffde0000 0x20000 0xffdff000 [name:iova&] 34 0xffffff80c8395380 0xffe00000 0x80000 0xffe7f000 [name:iova&] 35 0xffffff80c8395880 0xffec0000 0x20000 0xffedf000 [name:iova&] 36 0xffffff80c83957c0 0xffef9000 0x1000 0xffef9000 [name:iova&] 37 0xffffff80c83956c0 0xffefa000 0x1000 0xffefa000 [name:iova&] 38 0xffffff80c8395f40 0xffefb000 0x1000 0xffefb000 [name:iova&] 39 0xffffff80c8395a80 0xffefc000 0x4000 0xffeff000 [name:iova&] 40 0xffffff80c1b3e840 0xfff00000 0x100000 0xfffff000 [name:iova&] 41 0xffffff80d5be2860 0xfffffffffffff000 0x1000 0xfffffffffffff000 [name:iova&]dump_iova_rbtree done, iovad:0xffffff80d5be2808, count:41, total_size:614788KB

Thanks, Yunfei.